cepr6198textcrest.science/RePEc/wpstorage/2007-04.pdf · THE EFFECTS OF SEGREGATION AND SPATIAL MISMATCH ON UNEMPLOYMENT: EVIDENCE FROM FRANCE* Laurent GOBILLON** and Harris SELOD***

INSTITUT NATIONAL DE LA STATISTIQUE ET DES ETUDES ECONOMIQUES Série des Documents de Travail du CREST

(Centre de Recherche en Economie et Statistique)

n° 2007-04

The Effects of Segregation and Spatial Mismatch on Unemployment :

Evidence from France

L. GOBILLON 1

H. SELOD 2

Les documents de travail ne reflètent pas la position de l'INSEE et n'engagent que leurs auteurs. Working papers do not reflect the position of INSEE but only the views of the authors.

1 INED, 133 boulevard Davout, 75980 Paris Cedex 20, France.([email protected]) 2 INRA, Paris School of Economics, CREST and CEPR. Address : INRA-LEA, 48 boulevard Jourdan, 75014 Paris, France. ([email protected]) Corresponding address : INRA, 48 boulevard Jourdan, 75014 Paris, France.

THE EFFECTS OF SEGREGATION AND SPATIAL MISMATCH

ON UNEMPLOYMENT: EVIDENCE FROM FRANCE*

Laurent GOBILLON** and Harris SELOD***

January 13, 2007

Abstract

In this paper, we investigate how residential segregation and bad physical access to jobs contribute to urban

unemployment in the Paris region. We first survey the general mechanisms according to which residential segregation

and spatial mismatch can have adverse labor-market outcomes. We then discuss the extent of the problem with the help

of relevant descriptive statistics computed from the 1999 Census of the Population and from the 2000 General

Transport Survey. Finally, we estimate the effect of indices of segregation computed at the neighborhood and

municipality levels, as well as job accessibility indices on the labor-market transitions out of unemployment using the

1990-2002 Labor Force Survey. Our results show that neighborhood segregation is a key factor that prevents

unemployed workers from finding a job. These results are robust to potential location endogeneity biases.

Key words: residential segregation, spatial mismatch, urban unemployment, sensitivity analysis. JEL Classification: J64, R14.

Résumé

Ce travail étudie la façon dont la ségrégation résidentielle et la déconnexion physique aux lieux d’emploi

contribue au chômage urbain en Ile-de-France Nous rappelons dans une revue de la littérature les mécanismes généraux

selon lesquels la ségrégation résidentielle et le mauvais appariement spatial sont sources de difficultés sur le marché du

travail. Nous présentons ensuite des statistiques descriptives pour illustrer ces questions en Ile-de-France à partir du

Recensement de la Population de 1999 et de l’Enquête Globale des Transports en 2000. Enfin, nous estimons l’effet

d’indices de ségrégation calculés à l’échelle du voisinage et de la commune et d’indices d’accessibilité aux emplois sur

les transitions hors du chômage en utilisant les Enquêtes Emploi 1990-2002. Nos résultats montrent que la ségrégation

résidentielle est un facteur clé qui empêche le retour à l’emploi des chômeurs. Ce résultat est robustes aux biais

d’endogénéité de localisation.

Mots-clés: ségrégation résidentielle, spatial mismatch, chômage urbain, analyse de sensibilité. _______________________ * We thank Thierry Magnac, Eleonora Patacchini, Edmond Préteceille, Muriel Roger, Jos Van Ommeren and the participants to seminars at INRA and to the North American Meetings of the RSAI 2006 for useful comments. This project was carried out with the financial support of the Sustainable Urban Development program 2004-2006 (Ministry of Research, France). We are grateful to INSEE and the Ministry of Transport’s Regional Directorate (DREIF) for providing us with the data. We are responsible for any remaining errors in the paper. ** INED. Address: INED, 133 boulevard Davout, 75980 Paris Cedex 20, France. E-mail: [email protected] *** INRA, Paris School of Economics, CREST and CEPR. Address: INRA-LEA, 48 boulevard Jourdan, 75014 Paris, France. E-mail: [email protected] Corresponding address: INRA, 48 boulevard Jourdan, 75014 Paris, France.

1 Introduction

In november 2005, an isolated act of police violence that occurred in a low-income neighborhood out of Paris

triggered an unexpected wave of riots that quickly spread to many distressed areas across the country. Even

though suburban riots had sporadically taken place over the past three decades, most observers were surprised by

the renewed intensity of the phenomenon (9, 193 cars were burnt over a three-week period) and, for the first time,

by its generalization to more than 170 poor areas, most of them located in the suburbs of major metropolitan

areas. A parallel that comes to mind with this unprecedented phenomenon for France are the 1965 Watts riots

in the US. At that time, the Kerner Commission (1968) identified local unemployment as a major cause of social

unrest, and pointed at its potential determinants: geographic isolation and bad physical job accessibility. These

ideas have been much studied in the literature on the US over the past four decades (see the surveys of Jencks and

Mayer, 1990, Kain 1992, Ihlanfeldt and Sjoquist, 1998, and Gobillon, Selod and Zenou, 2007). In France, the same

type of explanations have recently been put forward by the Council of Economic Advisors (Fitoussi, Laurent, and

Maurice, 2004): workers residing in distressed and isolated areas simply do not have enough job opportunities.

To explore this idea, this paper estimates the effects of residential segregation and disconnection from jobs

on the labor-market transitions out of unemployment in the Paris region. The contributions of the paper to the

international literature on the topic are fourfold. First, we use the Labor Force Survey (LFS) that offers a unique

sampling scheme by exhaustively surveying all residents in very small neighborhoods (sometimes down to the

block level), which allows us to construct contextual indices at a very fine spatial scale. Existing studies on local

unemployment use data at a larger scale that is at best the census tract (Weinberg, 2000, 2004), one exception being

Bayer, Ross and Topa (2005) who use data at the block level. Second, since the LFS is panel, we are able to study

unemployment in a dynamic setting. This type of study is still quite scarce in the literature (exceptions being Holzer,

Ihlanfeldt and Sjoquist, 1994, Rogers, 1997, Dawkins, Shen and Sanchez, 2005). Third, we address the problem of

potential location endogeneity. This is done generalizing the existing literature on sensitivity analysis (Rosenbaum

and Rubin, 1983, Harding 2003) to multiple locations and multiple discrete outcome settings. Fourth, this paper

adds to the recent literature in Europe which is still emerging (see Fieldhouse, 1999, for the UK, Aslund, Osth and

Zenou, 2005, for Sweden, Dujardin, Selod and Thomas 2007, for Belgium, and Dujardin and Goffette-Nagot, 2005,

for France).

The structure of the paper is as follows: in a first section, we present the main theories according to which

residential segregation and physical disconnection from jobs can impede unemployed workers from finding a job

when they reside in disadvantaged areas. In a second section, we present stylized facts on the Paris region. We

show that, on average, local unemployment goes along with residential segregation but not necessarily with bad job

accessibility. In a third section, we present our microeconometric study of the effect of the local context on the labor-

market transitions out of unemployment. We show that unemployed workers residing in segregated neighborhoods

experience serious additional difficulties in finding a job. We find that this result is robust to biases induced by a

potential endogeneity of location. A fourth section concludes.

1

2 Residential segregation and access to jobs: a brief survey of the

literature

We present a selective review of the literature that shows how residential segregation and bad job accessibility can

deteriorate labor-market outcomes. We leave aside the origins of segregation and the physical disconnection from

jobs as this is not the focus of the paper (for details on the causes of segregation and disconnection from jobs, see

Gobillon et al., 2007).

2.1 The effects of residential segregation

A first trend of literature stresses the potentially harmful role of residential segregation on the economic outcomes

of poor-area residents. The main arguments are threefold: residential segregation deteriorates employability, it

reduces the quality of the social newtorks used in job-search activities, and it spurs labor-market discrimination.

This subsection briefly surveys the theories and intuitions that support these assumptions.

2.1.1 The low employability of segregated workers

The economic and sociological literature identifies two main channels according to which which residential segre-

gation can strongly deteriorate the employability of residents in poor areas.

The first channel involves the emergence and diffusion of social problems in segregated areas in line with the

“epidemic theory of ghettos” (Crane, 1991) which recognizes that social problems are endogenously determined

and that, once a given threshold of local social deterioration has been reached, the neighborhood is likely to face

a severe outburst of social problems. This occurs in the presence of a local externality when the propensity of a

young individual to adopt a deviant behavior (e.g. becoming pregnant, dropping out of school, or indulging in

criminal activites) is an increasing function of the proportion of same-behavior individuals in the neighborhood.

The local contagion effects that explain this pattern are well documented (Case and Katz, 1991). Their prevalence

is all the more increased as adults in these areas are themselves socially excluded and do not provide an image of

success with whom youngsters could identify (Wilson, 1987).

The second channel involves the difficulties faced by segregated residents to acquire human capital, a phenomenon

which finds two types of explanations. The first explanation is that segregation exerts a negative externality on the

learning process and deteriorates the quality of education acquired in schools (Benabou, 1993). Several studies,

indeed, confirm the existence of peer-group effects in the production of education and find that a child’s success

depends on the socio-economic characteristics of other students in the class or in the neighborhood (see e.g. Jencks

and Mayer, 1990, Ferguson, 1991, Hoxby, 2000). Focusing on the French case, Goux and Maurin (2004) show that

the local concentration of children with difficulties at school increases the individual risk of educational failure

–measured by the probability of repeating a grade–. Consequently, in areas where many students perform badly,

human capital externalities can further deteriorate the educational achievements of all the children in the area.

The second reason why segregation may cause bad educational outcomes is that it often generates low public

2

investments in educational inputs. In France, where education is overwhelmingly public, the Ministry of Education

is often blamed for placing young teachers with little or no work experience in poor areas while more experienced

teachers are not required to work in such places (Piketty, 2004). In France, the spatial concentration of poor

families also reduces local tax bases (Gilbert, 2004) and makes it difficult to finance key local public goods that

are needed for educational purposes such as libraries, stadiums or theatres.

2.1.2 The low quality of social networks in segregated neighborhoods

Another adverse effect of residential segregation is to deteriorate the quality of the social networks unemployed

workers can mobilize in order to find a job. This is a crucial problem for distressed areas given that a significant

portion of jobs is usually found through personal contacts (Mortensen and Vishwanath, 1994) especially among low-

skilled workers, young adults and ethnic minorities who largely resort to informal research methods (Holzer, 1987,

1988). In this respect, the spatial concentration of unemployed workers exerts a negative externality which amplifies

the probability of unemployment. In particular, a job seeker residing in a high-unemployment neighborhood knows

very few occupied workers –if any at all– that could personnaly support a job application with their employer or

that could gear them towards other professional contacts (Reingold, 1999, Selod and Zenou, 2001 and 2006).

2.1.3 Neighborhood stigmatization and redlining

A third link between segregation and bad labor-market outcomes involves a particular type of labor discrimination

based on an applicants’ place of residence. The main argument is that residential segregation –along a variety of

dimensions covering the spatial concentration of unemployed workers, ethnic minorities, low-skilled workers and

social housing– can result in the collective perception of “good” and “bad” neigborhoods and influence the hiring

behavior of employers. This stigmatization of neighborhoods is at the root of a discriminating behavior –known as

“redlining”– in which employers discriminate against the residents of distressed areas just as if their neighborhood

had been circled in red on a map.1

In theory, there are two principal motives that explain redlining. Firstly, redlining may reflect an employer’

subjective hostility or “taste for discrimination” (Becker, 1957) towards the residents of some neighborhoods,

rejecting their manners and habits (i.e. language, dress code, imputed social and religious behavior). This hostility

is sometimes grounded on the preferences of the customers themselves who could be reluctant to have contacts with

employees residing in stigmatized areas. Secondly, redlining may also occur in the absence of any prejudice when

employers have imperfect information on the candidates but use their neighborhood of residence as an indication

of productivity. This “statistical discrimination” (Phelps, 1972) is detrimental to distressed areas when employers

impute lower working abilities to their residents.

In France, redlining or a so-called “délit de sale adresse” is often denounced in the popular press but remains

1Another type of redlining that has labor-market consequences occurs on the credit market. All things else being equal, if loan

applications are more likely to be rejected when originating from distressed areas, then entrepreneurs who want to open a business or

to expand their activities may be impeded in their recruitment prospects.

3

empirically little studied. To our knowledge, the only work that tests the existence of redlining in France is Amadieu

(2004) which shows that same-skill candidates applying for a given job obtain less interviews when pretending to

reside in le Val Fourré –a highly stigmatized neighborhood in the suburbs of Paris– than when pretending to

reside in central Paris.2

2.2 The effects of the physical disconnection from jobs

A second trend of literature argues that the physical disconnection between places of residence and places of

work can have adverse labor-market outcomes, an intuition known as the “spatial mismatch hypothesis” in the

US context but which may also apply to other contexts. This assumption is supported by two different types of

arguments which respectively focus on job search and commuting costs.

2.2.1 Distance to jobs, job search and unemployment

Three different mechanisms argue that job search is inefficient, little intense and costly for the unemployed workers

who reside far away from job opportunities.

Firstly, the inefficiency of distant job search can be caused by informational frictions when the available infor-

mation about job vacancies decreases with distance to job opportunities (Rogers, 1997, Ihlanfeldt and Sjoquist,

1990, Ihlanfeldt, 1997, Wasmer and Zenou, 2002 and 2005). One reason is that many firms seeking to fill in a

vacancy for a low-skilled position mainly resort to recruitment methods with a limited spatial span such as posting

“wanted” signs in shops or having job openings advertised in local newspapers (Turner, 1997). Another reason

is that job seekers have difficulties identifying potential employers located in distant zones with which they are

unfamiliar. In this respect, Davis and Huff (1972) believe that job seekers only search efficiently within a restricted

perimeter around their residences even if this zone only hosts poorly remunerated low-skilled jobs.

Secondly, the low intensity of the job-search effort among job seekers distant from job opportunities can be

explained by an indirect argument based on housing prices: because accessibility is bad in places distant from jobs,

housing prices in these locations tend to be relatively lower, providing too little incentives for local job-seekers to

search actively. On the contrary, the unemployed workers who live close to jobs should be more impatient to find

a job in order to be able to pay their rents (see Smith and Zenou, 2003, for a model, and Patacchini and Zenou,

2006, for an empirical investigation).

Lastly, the cost of job search –which increases with distance– may limit the spatial span of job search. Indeed,

unemployed workers can give up searching far away from their residences when the costs associated with job search

become too large (Ortega, 2000).

2 In France, non-spatial forms of discrimination remain more studied (see e.g. Bataille, 1997). Recent trials suggest that sheer racial

or ethnic discrimination is also a widespread phenomenon.

4

2.2.2 Commuting and unemployment

The physical disconnection between places of residence and job locations imply long and costly commuting trips,

which may generate adverse labor-market outcomes.

The main argument stresses that unemployed workers must consider the offered wages net of transport costs

when deciding to accept or reject a job offer so that, if tranport costs are very large, workers are likely to reject

many job offers and remain unemployed or accept jobs that are poorly remunerated but located close to their

residences (Brueckner and Martin, 1997, Coulson, Laing and Wang, 2001, Brueckner and Zenou, 2003). This can

be a central issue for unskilled workers who do not own a car and who reside in areas where the public transport

system is of low quality (as a result of incomplete network coverage, long waits at interconnection nodes, or the

lack of coordination between transport modes).3 In particular, the unskilled workers eligible for part-time jobs

or who would be required to work very early in the morning or late at night may be confronted with low public

transport frequencies, if not with the closing of the network at the times when they would most need it.

A second argument that links commuting costs and unemployment has employers discriminate applicants who

live far away. This behavior can be understood if distance deteriorates productivity for instance because distant

workers are likely to come to work tired or arrive late because of transportation problems (Zenou, 2002).

3 Residential segregation and disconnection from jobs in the Paris

region

Keeping in mind the theories and intuitions presented in the previous section, we now discuss a selection of

relevant descriptive statistics for the year 1999. All the data we give are for the Paris region –that the French

call Ile-de-France–, an administrative unit of km2 12, 072 which hosts 10.9 million people distributed over 1, 300

municipalities centered around the historic city of Paris.4 The Paris region –which hosts 11.1 million people–

broadly corresponds to the statistical definition of the Paris urban area. The historic city of Paris itself groups 20 of

these munipalities (the so-called 20 arrondissements of Paris) and is home of 2.1 million residents. It is only a small

part of the metropolitan area’s urban core which also includes 395 surrounding municipalities and hosts a total

of 9.6 million residents. Most of the 885 remaining municipalities in the Paris region belong to the metropolitan

area’s periurban ring and a few of them are rural.5

3For a theoretical paper on public under-investment in transport systems, see Brueckner and Selod (2006).

4 In France, a municipality (commune in French) is the smallest autonomous administrative subdivision that exists. Urban areas

typically spread over several municipalities.

5A widely used administrative breakdown of the Paris region that we represent on our maps divides the whole region in 8 districts

(called “départements” in French). At the center of the region, the 20 municipalities of the historic city of Paris form one district.

Clockwise around Paris are the Hauts-de-Seine (west), the Seine-Saint-Denis (north east) and the Val-de-Marne (south east). The

combination of these three districts forms the inner ring (“petite couronne”) which more or less corresponds to the metropolitan area’s

urban core. The remaining peripheral districts are the Yvelines (west), the Val d’Oise (north), the Seine-et-Marne (east), and the

Essonne (south). Together, they form the outer ring (“grande couronne”) which is essentially made of periurban municipalities and of

5

The most striking fact are the large differences in local unemployment rates across municipalities and the high

concentrations of unemployment in just a few areas. Map 1 shows that local unemployment rates are much higher

in the north-east of Paris and in the municipalities located to the north and to the north east out of Paris (i.e. in

most of the municipalities in the Seine-Saint-Denis district and in some municipalities in the south of the Val-d’Oise

district).

[Insert Map 1]

As we are interested in how segregation and job accessibility relate to unemployment, we now present a selection

of descriptive statistics to assess the intensity of the two phenomena in the Paris region, and of their association

with local unemployment rates.

3.1 Segregation measures

We measure residential segregation along three dimensions: education, occupation, nationality.

We first measure the extent to which the residences of workers are spatially segregated along education levels.

Our measures of segregation include both local specialization indices and global dissimilarity indices. For a given

subgroup of the population and for each municipality, the local specialization index is given by the ratio of the

proportion of individuals residing in the considered municipality that belong to the subgroup to the proportion of

individuals in the whole region that belong to the same subgroup. When the specialization index takes a value

greater than one, the group is locally over-represented in the municipality. Map 2 graphs the specialization index

for educated workers in all municipalities and shows that they are over-represented in the city of Paris and in the

west of the Paris region (namely in municipalities in the Hauts-de-Seine and Yvelines districts). On the contrary,

low-educated workers are over-represented in the rest of the region. The segregation pattern is quite similar when

one considers occupation according to the French official job-positions scale that distinguishes high-skilled white

collars (“cadres”) from clerks (“employés”) and from blue collars (“ouvriers”). High-skilled white collars are

over-represented in the West. Blue collars are over-presented in the rest of the region. Clerks are frequently over-

represented in inner-ring municipalities (see Gobillon and Selod, 2004, for the corresponding maps).

[Insert Map 2]

The dissimilarity index (Duncan et Duncan, 1955) is of a different nature. It provides a global measure of segre-

gation between two types of occupations, yielding a single figure that measures segregation across all municipalities

between these two types. The dissimilarity index, which lies between 0 and 1, compares the spatial distributions of

two occupations by measuring the proportion of individuals from one of the groups that would have to be relocated

to other municipalities in order to have the two groups mixed in the same proportion in each municipality.6 Table

a few rural ones.

6The dissimilarity index is given by the formula 12

Pk

¯̄̄JikJi− Jjk

Jj

¯̄̄, where k is the municipality, Jik (resp. Jjk) is the number of

occupied workers with job occupation i (resp. j) residing in municipality k, Ji (resp. Jj) is the total number of occupied workers with

job occupation i (resp. j) in the whole Paris region.

6

1 presents the dissimilarity indices for our three combinations of occupations taken two by two. Observe that the

levels of segregation between high-skilled white collars and blue collars, and between high-skilled white collars and

clerks, are the highest: it would be necessary to relocate 39% (resp. 26%) of high-skilled white collars to other

municipalities in order to obtain a constant ratio of high-skilled white collars and blue collars (resp. clerks) in each

municipality. Also oberve that segregation is stronger in the inner ring than within Paris or across municipalities

within the outer ring.7

[Insert Table 1]

When computing segregation indices for nationalities, we chose to focus on Maghrebines and Africans (i.e.

Africans that are not Maghrebines) as these two groups form the bulk of disadvantaged minorities in France.8 Map

3 shows that both Maghrebines and Africans (excluding maghrebines) are concentrated to the north of Paris (in

the Seine-Saint-Denis district and the southern part of the Val-d’Oise district) as well as to the south of Paris (in

the western part of the Val de Marne district). Table 1 confirms that Maghrebines and Africans are significantly

segregated from the French since enforcing a uniform residential mix with the French would require to relocate

32% of individuals that belong to these two groups. Interestingly, Maghrebines and Africans are not distributed

very differently across space and reside in the same areas since the dissimilarity index between them is only .13.

[Insert Map 3]

Segregation by diploma is not as stark as segregation by job occupation or by nationality. Only 25% of the

educated labor force (i.e. workers holding at least a high school diploma) would have to be relocated to other

municipalities if one wanted to obtain a uniform residential mix of educated and uneducated workers in each

municipality (see Table 1).

Considering all these dimensions (occupations, nationalities, and diplomas), Paris appears to be a more inte-

grated area than the inner and outer suburban rings as segregation indices are always lower within Paris (i.e. across

Paris districts).

Now, having identified different types of residential segregation, we can measure the association between seg-

regation and unemployment in the Paris region. We find that the unemployment rate is strongly correlated with

the local composition of each municipality. Indeed, the correlation between the unemployment rate and the pro-

portion of high-skilled white collars is −.54. The correlation between the municipality unemployment rate and theproportion of Africans (including Maghrebines) stands at .85.9

7For comparaison, Cutler, Glaeser and Vigdor (1999) who measure black/white segregation in american MSAs consider that segre-

gation is low when the dissimilarity index is below .3, that is is medium when the index is between .3 and .6, and that it is high when

the index is above .6. With this grid of analysis, our measures would indicate only a low or medium level of segregation. However,

Cutler et al. use a finer spatial breakdown (the census tract) than we do, and it is well-known that segregation indices usually decrease

with the size of the geographic units considered. By comparing segregation indices at different spatial levels, Préteceille (2003) checked

that this decreasing pattern is also the case in the Paris region.

8 In France, the collect and use of racial or ethnic statistics are forbidden by law so we are constrained to work on nationalities.

9All the correlations in this subsection and in the the following one are computed weighting the observations by the labor force in

7

3.2 Job-accessibility measures

We measure the disconnection between places of residence and job locations in two different ways, with the calcu-

lation of global dissimilarity indices and local ratios of labor demand and supply.

The dissimilarity indices we compute compare the spatial distribution of workers and that of jobs. In Table

2, we report the indices for all workers and all jobs but also by occupation type, nationality and diploma.10 . The

dissimilarity index between all workers and jobs is .25, but this figure conceals differences by occupation type. For

high-skilled white collars and their corresponding jobs, the index reaches .31, whereas for clerks, it stands at .30,

and for blue collars, it is .25. Observe that blue collars exhibit the lowest disconnection from jobs because of the

decentralization of manufacturing firms to the inner and outer rings where blue-collar workers traditionally reside

–and where land is cheaper and space readily available–. The dissimilarity indices by nationality also reveal

stark differences. With a value of .33, it is highest for Africans (excluding Maghrebines) and the jobs they occupy,

but stands only at .26 for Maghrebines, very close to the value of the index for the French (.25). Interestingly,

focusing on dissimilarity indices by educational level, we obtain the same value of the index (.26) irrespectively of

whether we focus on educated workers and jobs (i.e. with a high school diploma) or on uneducated workers and

the corresponding jobs.

[Insert Table 2]

A second approach measures the ratio of available jobs to resident workers (whether occupied or not) within

a 45-minute commuting trip around each municipality. This is the most simple way of capturing the tension on a

local labor market centered on each municipality. In Appendix A, we present two alternative indicators based on

refined assumptions on the local labor demand and supply. The reason we have chosen a 45-minute cut-off is that

it is slightly above the average commuting time to work in the Paris region (which stands at 34 minutes in 1997

according to DREIF-INSEE, 1997). We believe it approaches the relevant area within which a worker would accept

a job without a residential move. The indicator is computed using job and worker locations from the 1999 Census

of the Population and the commuting times between municipalities estimated at morning peak hours in 2000 by

the Ministry of Transport for both public and private transport modes.11

Job densities within 45 minutes using public tranport are represented on Map 4. It can be seen that job

accessibility by public transport is very good in Paris, in the north of the Hauts-de-Seine district, in the west of

the Seine-Saint-Denis district and in the west of the Val-de-Marne district. In these areas, each municipality is

the center of an isochron inside which there are 1.2 to 2 times more jobs than workers. Map 5 shows that job

each municipality.

10For occupation types, dissimilarity indices are computed on occupied workers and the corresponding jobs. This is because, by

definition, unemployed workers have no job occupation. By contrast, when computed for a particular educational level or nationality,

our dissimilarity indices measure the disconnection between all workers of the considered type (whether occupied or not) and the

corresponding jobs.

11Matrices of commuting time between municipalities were estimated by the Ministry of Transport (Direction Regionale de

l’Equipement Ile-de-France) using the General Transport Survey (Enquête Globale des Transports, 2000). For a detailed description of

this data and their use to calculate other accessibility indices, see Wenglenski (2003).

8

accessibility is on average lower by private vehicle than with public transport (because of freeway congestion). Job

accessibility is better in and around Paris (except to the east) and to the west, including in the outter ring. To

summarize the information given by the maps, we also computed the average density reachable in 45 minutes for

all the municipalities in the Paris region, in the inner and outer rings, and in the historic city of Paris (weighting

by the local labor force). Table 2 confirms that access to jobs is on average better by public transport than with

private vehicles, and that it decreases when moving away from Paris. We also computed job-accessibility indices

separately for uneducated and educated workers (i.e. without and with a high school diploma). Job accessibility

by private vehicle is not as good for uneducated workers than for educated ones. Conversely, job accessibility by

public transport is better for uneducated workers than for educated workers in Paris and in the inner ring.

[Insert Maps 4 and 5]

We finaly examine the correlations between unemployment and our measure of access to jobs. The unemployment

rate is not correlated with the density of jobs that can be reached within 45 minutes by private vehicle (the

correlation being only .01) but it is positively though weakly correlated with the density of jobs that can be reach

by public transport within the same time span (with a correlation of .15). A priori, this suggests that unemployed

workers may not suffer from a poor access to jobs in the Paris region.

4 The econometric setting

The stylized facts we just presented showed that there are large unemployment disparities among the region’s

municipalities. Areas where the unemployment rate is high usually exhibit a high level of segregation but not

necessarily a spatial disconnection from jobs. Using individal data, we now study how the local context (segregation

and job accessibility) can affect the labor-market transitions out of unemployment. Studying transitions rather

than the cross-section labor-market status (i.e. being unemployed or occupied at one point in time) as in Dujardin,

Selod and Thomas (2007) makes the interpretation of the results easier because the job-search mechanisms that

we described in Section 2 are dynamic by definition.

4.1 The data

The dataset

Studying the effect of the local context on the transitions out of unemployment requires panel data with infor-

mation on the place of residence at a fine geographical scale. To our knowledege, the French Labour Force Survey

(LFS) is the only data source that can be used for this purpose. This dataset is a rotating panel where individuals

are surveyed yearly over a three-year period but only if they do not move. If they move, they disappear from the

panel. The sampling scheme consists in surveying tiny neighborhoods of nearly twenty households living in adjacent

dwellings (see Goux and Morin, 2004 for a more detailed description of the sampling scheme). All individuals older

than 15 are surveyed. The dataset contains the usual socio-demographic information (sex, age, diploma) as well

9

as variables on unemployment and employment spells. We selected only workers that are unemployed (according

to the definition of the International Labour Organization) and who reside in the Paris region.12 We then con-

structed the yearly transitions over the 1990-2002 period that correspond either to a labor-market transition or to

an exit from the dataset that occurred too early (because of a move, death, non-response, etc.). We obtained a

final subsample of 9, 643 transitions of four types: remaining unemployed (33.7%), finding a job (26.2%), leaving

the labor force (13.9%) and disappearing from the dataset too early (26.2%). The first three types of transition

are implicitely not associated with any residential move. To the contrary, the fourth type of transition mainly

corresponds to residential moves.13 The location of unemployed workers experiencing a transition are distributed

across 3, 905 neighborhoods-years; the surveyed neighborhoods being located within 371 muncipalities of the Paris

region, mainly in the urban core, i.e. in the hitoric city of Paris and in the inner ring. Map 6 represents the number

of transitions in each muncipality over the whole period.

[Insert Map 6]

The local context

Thanks to the spatial cluster structure of the LFS survey, we are able to compute contextual measures of the

immediate neighborhood with which unemployed workers are likely to interact when looking for a job.14 For each

neighborhood, we computed the proportion of high-skilled white collars, the proportion of educated workers in the

labor force (i.e. having a high school diploma), the rate of Africans (including Maghrebines), and the unemployment

rate. It would have been desirable to also compute a job accessibility measure at the neighborhood level but data

required for this are not available. Descriptive statistics at the neighborhood-year level are reported in Table 3.

The number of observations by neighborhood-year varies from 1 to 109, with an average number of 27 observations.

[Insert Table 3]

We also constructed the same proportions at the municipality level using the data from the 1999 census. We

report in Table 4 the correlations between the indices computed at the neighborhood-year level for the 1990-2001

period and at the municipality level in 1999. All the correlations are computed weighting the observations by

the number of transitions observed for each neighborhood-year. They characterize the local environment of the

surveyed unemployed workers in the LFS at two different spatial scales.

At the municipality level, the proportion of high-skilled white collars is negatively correlated with the proportion

of Africans (-.47). This means that when municipalities where unemployed workers live exhibit a large proportion of

12According to the ILO definition, an unemployed worker is an individual old enough to work (i.e. over fifteen years of age), available

to work and looking for a job.

13Using the French LFS, Desplanques (1994) estimates that, for women aged between 20 and 64, 85% of the exits from the dataset

that occur too early are due to a residential move.

14Héran (1987) shows that the spatial span with which individuals declare to have social interactions decreases with population

density. In Paris for instance, 80% of the households declare that they only interact with neighbors in their building. It is likely that

the spatial span of the French LFS’s clusters, which is small and also diminishes with density, approximates the relevant neighborhood

scale for social interactions.

10

high-skilled white collars, they tend to have a small proportion of Africans. The correlation between the proportion

of high-skilled white collars and our index of job accessibility by private transport are strongly positively correlated

(0.60). By contrat, the correlation between the proportion of Africans (including Maghrebines) and the density of

jobs that can be reached using a private vehicle is lower (.17). The correlation between the percentage of high-

skilled white collars and the job density that can be reached by public transport within 45 minutes is .32. The

correlation between the percentage of Africans (including Maghrebines) and the job density that can be reached

by public transport is also .32.

At the neighborhood scale, the correlation between the proportion of high-skilled white collars and the pro-

portion of Africans is −.35. Moreover, neighborhood variables are strongly correlated with variables of the sametype at the municipality level. In particular, the correlation between the proportion of neighbors that are high-

skilled white collars and the proportion of high-skilled white collars in the municipality is .60. The correlation for

the neighborhood and municipality proportions of Africans (including Maghrebines) is .48. These correlations are

quite high but still suggests some heterogeneity between neighborhoods in a same municipality.

[Insert Table 4]

To better characterize neighborhoods, we classify them along the following three dimensions: neighborhood

segregation, municipality segregation and job accessibility. To do that, we run a separate hierarchical ascending

classification (HAC) for each dimension keeping only two classes each time. For neighborhood segregation, the

HAC is built from the neighborhood proportion of educated workers and that of Africans. The two classes oppose

segregated neigborhoods to non-segregated neighborhoods (see Table 5a). For segregation at the municipality level,

the HAC is built from the municipality proportion of educated workers and that of Africans. The two classes oppose

neighborhoods in segregated municipalities to neighborhoods in non-segregated municipalities. For job accessibility,

the HAC is built from job densities by private and public transport. The two classes oppose neighborhoods in

municipalities with good job acessibility to neighborhoods in municipalities with bad job acessibility. Crossing

the three HAC, we construct an eight-class typology of neighborhoods which is described in Table 5b. In every

municipality, we find that at least 50% of the LFS transitions occur in a given neighborhood type. Map 7 represents

the prominent type across municipalities. Consistenly with the stylized facts presented in the previous section, we

see an opposition in the degree of neighborhood and municipality levels of segregation and between the west and

the north-east of Paris and its inner ring. The north-east where segregation is strong is split between municipalities

close to Paris that have a good job-accessibility and those further away that have a bad job accessibility.

[Insert Tables 5a and 5b]

[Insert Map 7]

4.2 The model

It would be natural to study unemployment in a dynamic perspective using an unemployment duration model

but our data provides only incomplete coverage of unemployment spells. In this context, estimating a duration

11

model would be feasible but very burdensome and in any case not completely satisfactory as duration dependence

would not be considered (see Magnac, Robin and Visser, 1995). Instead, we study the labor-market transitions of

unemployed workers with a discrete-choice model. It would be tempting to resort to panel data techniques (with

individual random or fixed effects) but this would require to observe mutliple transitions for each type of exit out

of unemployment. This is not possible for transitions from unemployment to employment or inactivity which can

be observed at most once with three consecutive yearly observations. We thus simply use a multinomial model

with latent variables corresponding to E = 4 types of exits: remaining unemployed (e = 1), finding a job (e = 2),

dropping out of the labor force (e = 3), and disappearing from the dataset (e = 4). For a worker i at time t (as

each individual can experience up to two transitions over the period), the exit is the result of a random variable

denoted Eit. The latent variable associated to any exit e writes:

E∗it,e = Xitαe + Zjtβe + εit,e (1)

where Xit is the set of individual explanatory variables at time t (assumed to be exogenous), j is the worker’s LFS

neighborhood at the beginning of the transition. Observe that, since movers are not followed, when an individual

experiences two transitions (i.e. appears at three consecutive dates in the panel), his location does not change.

Zjt is the set of aggregate explanatory variables for location j at time t, and εit,e is a random shock. Here, j

can be viewed as a draw of a random variable Li that accounts for the location choice of a worker i. As we are

interested in measuring the effect of the explanatory variables Xit and Zjt on the labor-market transitions out of

unemployment, we choose exit e = 1 as a reference and impose the restrictions α1 = 0 and β1 = 0. We also assume

that the random shock εit,e is independent from the observable characteristics Xit and Zjt.

Assume for the time being that the location Li is independent from the random shock εit,e. The location choice

is then exogenous. We choose to specify εit,e as following an extreme-value law (conditionally on the observables)

so that we finally obtain a multinomial logit for the different exits where the probability that the individual i

experiences an exit Eit of type e ∈ {2, 3, 4} writes:

P (Eit = e |Xit, Li = j, Zjt ) =exp (Xitαe + Zjtβe)

1 +X

e=1,2,3

exp (Xitαe + Zjtβe)(2)

We estimate the model at two different geographic scales. In the first specification, the geographic scale is the

neighborhood (J = 2, 424). In this case, the variables Zjt include our job-accessibility and segregation variables

both at the neighborhood and the municipality levels. In the second specification, the geographic scale is the HAC

cluster (J = 8). In this case, the Zjt variables only include seven location dummies (one cluster being chosen as a

reference).

4.3 The sensitivity analysis

A major issue is that the estimates of (2) can suffer from a selection bias if unemployed workers sort themselves

spatially along some unobservables. More precisely, spatial sorting appears when there are individual unobserved

12

factors that simultaneously have an effect on the labor-market transition and the location choice. Put differently,

the location Li may be correlated with the random shock εit,e and the location is then endogenous. An approach

to overcome this problem could be to estimate an econometric model that would simultaneously explain the

transition on the labor market and the location choice. This model should allow for a correlation between the

individual unobservables introduced in the transition and location specifications. Nevertheless, the estimation of

such a model would be problematic for two reasons. First, a method for estimating a model with multiple outcomes

and multiple locations has only been proposed in the case where the outcome is continuous (Dahl, 2002). Second,

even if such a method were available, identifying the model without relying on non-linearities would require an

exclusion restriction such that some variables explaining location have no direct effect on labor-market transitions.

Finding such variables is very difficult. Given these problems, we choose to use two alternative strategies to address

the endogenity of location:

The first strategy follows Goux and Maurin (2004) and checks that our results are robust when reestimating

the model only on the subsample of unemployed workers living in public housing. Indeed, one can think that

their location is exogenous given that offers for a public dwellings are made by public-housing offices without

taking into account the preferences of applicants for locations. Two criticisms however should be kept in mind.

First, it is possible that some individuals reject a housing offer and prefer to wait for another one that better fits

their preferences (Dujardin and Goffette-Nagot, 2005). The endogeneity problem then still occurs, for instance

if individuals with valuable unobservable characteristics (i.e. that improve their chances to find a job), more

frequently reject housing proposals for suburban public dwellings where finding a job is more difficult. Second,

unemployed workers living in public dwellings may have unobservable characteristics that make it more difficult to

find a job. Hence, considering only individuals in public housing introduces a selection bias in the analysis.

The second strategy is a sensitivity analysis which goal is to evaluate the endogeneity bias created by an

unobserved individual variable that would affect both locations and transitions. We build on Rosenbaum and Rubin

(1983)’s approach for a binary logit model which was applied to test the robustness of local effects by Harding (2003)

and Dujardin, Selod and Thomas (2007). In this type of method, the key idea is to jointly model the transition and

location choices. The location endogeneity problem is accounted for by introducing a same unobserved individual

variable both in the transition model and the location choice model. The law of this unobserved variable may

be specified in different ways: it can be continuous or discrete–multinomial or, in particular, binary– and is

considered as known. Its effects in both the transition and location specifications are drawn in a given range of

values. The effects of explanatory variables in the transition equations (and specifically those of local variables)

are then reestimated. We assess whether these reestimated effects differ from those obtained with the simple

multinomial logit model without individual unobserved heterogeneity.

In the transition equation (1), for each exit e, the shock in the labor-market transition is split between the effet

of the unobserved individual variable ui and a random component ηit,e:

εit,e = γeui + ηit,e (3)

where γe measures the effect of the unobserved variable on the transition and ηit,e follows an extreme-value law.

13

The location choice among possible locations is assumed to follow a logit specification conditionally on the

unobserved and observed individual variables. In this model, denoting t0 the individual’s initial transition date,

the latent-variable equation for location j ∈ {1, ..., J} is specified as:

L∗i,j = Ait0λj +Bjt0θ + δjui + νi,j (4)

where Ait0 is a set of individual observed characteristics at date t0 with effects λj , Bjt0 is a set of observed aggregate

characteristics at date t0 with effects θ, δj is the effect of the unobserved individual variable and νi,j follows an

extreme-value law. Observe that the effects of the observed and unobserved indivividual variables (λj and δj

respectively) can depend on location.

The complete model finally includes equations (1), (3) and (4). The endogeneity issue is due to the fact that

the unobserved individual variable has an effect simultaneously in the labor-market transition equations (1) via (3)

and in the location-choice equations (4). It appears only if the coefficients γe and δj are simultaneously different

from zero and if there is some variation in the value taken by the individual unobserved variable among unemployed

workers. The analysis consists in estimating the effects of individual and agregate explanatory variables (αe and

βe) in (1) under alternative assumptions on the effects of the unobserved individual variable (γe and δj). Observe

that, since the unobserved heterogeneity coefficients are fixed –and not estimated– no exclusion restriction is

needed (Altonji, Elder and Taber, 2005). There is no need to have a variable in Ait0 or Bjt0 that is not in Xit or

Zjt.

The procedure can be decomposed into two successive steps:

• We draw the coefficients of the individual unobserved variable in the transition specification (γe) and in thelocation specification (δj) in some given specified laws.

• We reestimate the model by full maximum likelihood. This finally gives the effects of the explanatory variablesin the transition equations (1).

The methodological details are relegated in Appendix B. In our application, we carry out the sensitivity analysis only

for the transition specification with cluster dummies (J = 8). This is because performing the sensitivity analysis on

J = 2, 424 locations would have required a much greater number of transitions than what our sample provided. The

objective being to assess how the coefficients in the transition equations vary under different endogeneity patterns,

we repeat the procedure S = 1, 000 times. We report some descriptive statistics on the coefficients obtained from

the different simulations. In particular, we examine to which extent the coefficients obtained in our simulations are

of the same sign and significance as those obtained from the estimation of (2) when endogeneity is not accounted

for. The application details are presented in appendix B.

14

4.4 The results

4.4.1 A regression with segregation and job-accessibility indices

We first present the results when estimating the multinomial logit model (2). The reference category for transi-

tions is “remaining unemployed and not moving”. The regression includes sociodemographic individual variables,

variables related to job search, as well as the local indices measuring segregation and job accessibility. Since seg-

regation indices are highly correlated at the neighborhood and municipality level (see Table 4), we retain in the

main regression only two segregation indices which have a significant effect while being little correlated between

themselves. One of the indicator is measured at the neighborhood level in order to capture the effects that occur

at a fine geographic scale (such as interactions between neighbors). The other indicator is measured at the munic-

ipality level in order to capture the effects that occur at a broader geographic scale (such as redlining). This also

allows to control for the municipality’s composition while estimating the effects of the job accessibility variables.

The contextual variables that we retain are: concerning segregation, the percentage of neighbors in the active

population that have a high school diploma and the percentage of Africans in the municipality; concerning job

accessibility, the job densities within a 45 minute isochron by private transport and by public transport.

In Table 6a, we report the estimated coefficients and the corresponding odds-ratios. We first comment the

effects of individual characteristics on the different transitions.

Technical and university diplomas increase the probability of finding a job while having a general high school

diploma but no other diploma reduces it. Nationality also plays a role: being Maghrebine or African (but not

Maghrebine) diminishes the chances of getting a job (at the 10% level). Searching for a job through personal

methods in addition to being registered with the National Employment Agency increases the probability of finding

a job. Interestingly, searching for a job without registering with the National Employment Agency is more fruitful

than registering with the National Employment Agency and not resorting to any other job search method. This

can be understood if the choice of the search method is selective: those who might easily find a job could be less

likely to register with the National Employment Agency.

Concerning “dropping out of the labor force without moving”, what is striking is the negative effect of resorting to

personal contacts for individuals registered with the National Employment Agency. An explanation could be that

unemployed workers that are sufficiently motivated to diversify their job-search methods are less easily discouraged

and less likely to drop out of the labor force.

Finally, the category “disappearing from the panel” (an exit which is usually associated with residential mobility)

is favored by having a university diploma (at the 1% level), which is consistent with a greater residential mobility

of the educated, as previously observed in the literature (Schwartz, 1973). In addition, home owners have a lower

relative risk of disappearing prematurely from the panel than renters in the public sector, who themselves have a

lower relative risk of disappearing prematurely from the panel than renters from the private sector. These results are

comparable to those obtained on residential mobility by Gobillon (2001) on the European Community Household

Panel (ECHP).

We now comment the effects of the contextual variables on the three types of transition.

15

Concerning finding a job, the percentage of educated neighbors (i.e. that have a high school diploma or above)

has a positive effect, which is in accordance with social network and neighborhood stigmatization theories. The

percentage of Africans in the municipality has a negative effect (although not significant), which is consistent with

redlining. Among job-accessibility variables, only the job density within 45 minutes by public transport has a

significant effect (at the 10% level) but the impact is negative, which is not consistent with the spatial mismatch

theory. This result remains even when looking at jobs accessibility by skill level (see Appendix C). This raises

a series of comments. Firstly, problems of job accessibility may not impede unemployed workers from finding

a job. This was already suggested by Maps 1 and 4 in which job accessibility is good in municipalities where

unemployment rates are high. Secondly, the sign of the effect can be explained by the fact that job densities

may capture unobserved individual or local variables detrimental to finding a job. For instance, it is possible that

individuals with adverse unobserved characteristics have no other choice than to locate in unattractive suburbs

but with a good job accessibility. Conversely, individuals with beneficial unobserved characteristics may locate in

attractive suburban residential neighborhoods but far away from jobs. An unobserved local variable detrimental

to finding a job but correlated with job density is the congestion of public services (a well-known phenomenon in

France). Thirdly, our job density indicators are computed using transport times at the municipality scale, which

might occult relevant intra-municipality differences in job-access (for instance when an area is crossed by train

tracks but has no train station).

Concerning the transition out of the labor force, none of the contextual variables has a significant effect.

Concerning the premature exit from the panel, the proportion of educated workers in the neighborhood has a

positive effect while the proportion of Africans in the municipality has a negative effect (although only significant

at 10%). This is consistent with an increased mobility of the individuals that reside in neighborhoods populated

with educated workers and in municipalities with a low concentration of Africans.

[Insert Table 6a]

To test the robustness of the results of our baseline model, we also estimate a series of alternative specifications.

Since we use a multinomial logit model, transitions are assumed to verify the independence of irrelevant alternatives.

As this may not be the case in practice, we performed a robustness check by estimating a nested logit model where

the ιia is not a priori assumed. The objective is to check that the sign and significance of the effects of local

variables is comparable to what we obtained estimating (2). At the first level of the nested-logit model, an

individual disappears or stays in the panel (e = 0 vs e ∈ {1, 2, 3}). At the second level, conditionally on staying,the individual experiences a labor-market transition that can be: remaining unemployed (e = 1), finding a job

(e = 2) or dropping out of the labor force (e = 3). This specification allows for a correlation between the transitions

e = 1, 2 and 3. We obtain results that are very close to those of our baseline model. In particular, the neighborhood

proportion of high educated workers has a positive and significant effect on finding a job and disappearing from

the panel.15

15Details are available upon request.

16

For the baseline multionial logit model (2), we also ran a regression discretizing segregation and job-accessibility vari-

ables in order to capture possible non-linear effects. We tried alternative indices of segregation and job-accessibity

measures. For job-accessibility, we used indicators computed by level of education. None of these specifications

qualitatively altered our findings (for details, see Appendix C).

We finally examined whether the results could be generated by measurement errors due to the small size of some

neighborhoods, which would affect the percentages computed at the neighborhood level. This is done by excluding

observations that correspond to a neighborhood-year with only a small number of workers. We obtained results

similar to those in the main regression.16

4.4.2 A regression with dummies for neighborhood type

We also run a regression substituting dummies for the type of neighborhoods (see the previous section for their

construction) to the continuous contextual variables. The reference group is the class of neighborhoods with weak

segregation both at the neighborhood and municipality levels and a good job accessibility (type I). The results

are presented in Table 6b. We find that residing in a neighborhood that differs from the reference group only

because it is segregated (type III) reduces the chances of finding a job. Residing in a neighborhood that differs

from the reference group only because it has a bad job accessibility (type V) increases the chances of finding a

job, of dropping out of the labor force, and of moving (10% level only). This result does not support the spatial

mismatch hypothesis for this type of neighborhood, confirming our previous evidence. Residing in a neighborhood

that is segregated and located in a segregated municipality and has a bad job accessibility (type VIII), decreases

the chances of moving. The effect of living in such a neighborhood on finding a job, although of the expected

negative sign, is not significant.

[Insert Table 6b]

4.4.3 Accounting for a potential endogeneity of location

We now present the results of our two robustness tests to address the issue of location endogeneity. First, we

reestimate our model using the subsample of unemployed workers residing in social dwellings (3, 108 transitions for

which location “choice” is likely to be exogeneous).

We present the results for the regression with segregation and job-accessibility indices in Table 7a. They are

not very different from those obtained for the sample of all unemployed workers (Table 6a). The rate of educated

neighbors still has a positive and significant effect on finding a job and on the premature exit from the panel.

16When we exclude from the sample the observations which neighborhood-year has a number of workers below the first decile (which

corresponds to 12 workers), results are similar to those obtained in the main regression. The only noticeable difference is that the

negative effect of job density by public transport on finding a job is now significant at the 5% level. When we exclude the observations

which neighborhood-year has a number of workers below the median (which corresponds to 21 workers), results are once again very

close to those obtained in the main regression. However, the positive effect of the percentage of educated workers on finding a job is

not significant at the 10% level anymore.

17

The rate of Africans in the municipality has a negative effect on the premature exit from the panel and it is now

significant at the 1% level. This suggests that unemployed workers residing in municipalities where the percentage

of Africans is high are often less mobile. The density of jobs that can be reached via public transport still has a

negative effect on finding a job, but this effect is not significant. Finally, it is noticeable that the job density that

can be reached by private transport now has a positive effect on exiting the panel (significant at the 5% level) but

no significant effect on finding a job. This result suggests that when locations are exogenous, job accessibility via

private transport may positively influence mobility. Two interpretations seem possible. First, in accordance with

the spatial mismatch theory, the unemployed workers who reside in areas where job accessibility is good could have

better chances of finding a job, which could involve moving closer to a new place of employment. Alternatively,

the local job density could be correlated with unobserved individual or local factors having a positive influence on

residential mobility.

The results for the regression that uses dummies for neighborhood type are presented in Table 7b. In comparison

with the regression on the whole sample (Table 6b), most coefficients become not significant and even sometimes

change sign (without being significant). In particular, this occurs for neighborhoods with a weak segregation and

a bad job-accessibility (type V). Note also that the effect of living in a segregated neighborhood with a good job

accessibility (type III) still has a negative effect on finding a job but the effect is not significant anymore. Given

this accuracy problem, it is hard to conclude from this regression with neighborhood dummies on the subsample of

unemployed workers residing in public housing whether there is a location endogeneity issue or not. The accuracy

problem may stem from the reduction of the sample size and/or from a lack of variability as we aggregated

information through the HAC.

[Insert Tables 7a and 7b]

Our second approach is more elaborate and consists in evaluating a potential endogeneity bias due to an

unobserved individual variable that would simultaneously influence the exit from unemployment and the location

choice (see section 4.3 and Appendix B). Since carrying out the sensitivity analysis for the model with 2, 424

locations (neighborhoods) would have required many more observations, we focus only on the specification with 8

locations (clusters). We first consider cases in which there is at best a significant but moderate degree of endogeneity,

and simulate 1, 000 settings for the effects of individual unobserved heterogeneity .17 Our results (see Appendix

B) show that the coefficients of cluster dummies that were obtained for the multinomial logit regression (2) are

usually robust when the endogeneity of location is taken into account. The effect on finding a job of living in a

type-III neighborhood (neighborhood segregation) is negative in all the cases and significantly negative in 86% of

the cases. The effect on finding a job of living in a type-V neighborhood (bad job accessibility) is positive in all

17 Indeed, we draw the coefficients of the unobserved individual variable, γe and δj in [− ln 2, ln 2] such that the corresponding odds-ratios are comprised between .5 and 2. For comparison, observe that in our estimation of the baseline transition equations, odds-ratios

are comprised between .2 and 2.1 (see Table 6b). In the location-choice specification, they are comprised between .1 and 20.2 (see

Table A1).

18

the cases and significantly positive in 98% of these cases. This suggests that the negative effect of job accessibility

is robust when location endogeneity is moderate.

To check what happens in more problematic cases, we replicate the methodology three more times with increas-

ing degrees of endogeneity.18 Results are easily represented graphically. For instance, for cluster III (neighborhood

segregation), Graph 1 plots the percentage of estimated coefficients in labor-market transition equations that are

positive and significant (at a 5% level), positive but not significant, negative but not significant, and negative

and significant. As expected, for finding a job, the percentage of negative coefficients decreases with increasing

endogeneity patterns from 100% to 93%. The percentage of significantly negative coefficients decreases from 86%

to 55%. Graph 2 plots the same statistics for the effect of living in cluster V (bad job accessibility) on labor-market

transitions. When considering finding a job, the percentage of positive coefficients decreases from 100% to 93%

when endogeneity increases. The percentage of significantly positive coefficients also decreases from 98% to 70%.

This suggest that our result on neighborhood dummies are fairly robust to location endogeneity issues.

[Insert Graphs 1 and 2]

5 Conclusion

This paper studies the local determinants of urban unemployment in the Paris region. We present some key stylized

facts which characterize the intensity of segregation in the Paris region and differences in job accessibility across

the region’s municipalities: the locations where the unemployment rate is the highest are also characterized by

segregation but apparently not by a bad job accessibility.

We then estimate the effect of the local context on the individual transitions out of unemployment. Our analysis is

dynamic and adds to the literature that is mostly cross-section. Another originality of our paper is to use a dataset

with a unique sampling scheme that exhaustively surveys all residents in very small neighborhoods. This allows us

to construct neighborhood segregation indices at a much finer scale than most previous studies. Our results suggest

that neighborhood segregation prevents unemployed workers from finding a job. This could be explained by the

low quality of social networks in the neighborhoods where unemployed workers live or by a possible phenomenon

of redlining.

We test the robustness of this result to a potential location-endogeneity bias in two different manners. A first ap-

proach is to restrict the sample to public housing residents whose location can reasonably assumed to be exogeneous

in the case of France. A second approach resorts to a sensitivity analysis that generalizes the existing literature to

a setting with multiple discrete outcomes and multiple locations. We find that the negative effect of segregation on

finding a job remains.

18Technically, we draw 1, 000 simulations such that the odds-ratios for γe and δj are comprised between 1/3 and 3, another 1, 000

simulations such that they are comprised between 1/4 and 4, and yet another 1, 000 bewteen 1/5 and 5.

19

Ours results also suggest that segregation has an effect on residential mobility. An interesting issue in this respect

is whether residential mobility is associated with an enhanced return to employment. Since our dataset does not

offer this possibility, we leave this issue for future research.

20

6 References

References

[1] Altonji J., Elder T. and C. Taber (2005) Selection on observed and unobserved variables: assessing the

effectiveness of Catholic schools, Journal of Political Economy, 113(1), 151-184.

[2] Amadieu J.-F. (2004) Enquête testing sur CV, Adia/Paris 1 - Observatoire des discriminations, mimeo.

[3] Aslund O., Osth J. and Y. Zenou (2005) How crucial is distance to jobs for Ethnic Minorities? old question -

improved answer”, mimeo.

[4] Bataille P. (1997) Le racisme au travail, Paris, La Découverte, 266 p.

[5] Becker G. (1957) The Economics of Discrimination, Chicago: Chicago University Press.

[6] Benabou R. (1993) Workings of a city: location, education, and production, Quarterly Journal of Economic,

108, 619-52.

[7] Brueckner, J. and R. Martin (1997) Spatial mismatch: an equilibrium analysis, Regional Science and Urban

Economics, 27, 693-714.

[8] Brueckner J. and H. Selod (2006) The political economy of urban transport-system hoice, Journal of Public

Economics, 90, 983-1005.

[9] Brueckner, J. and Y. Zenou (2003) Space and unemployment: The labor-market effects of spatial mismatch,

Journal of Labor Economics, 21, 242-266.

[10] Case A. and L. Katz (1991) The company you keep: the effects of family and neighborhood on disadvantaged

families, NBER working paper 3705.

[11] Coulson E., Laing D. and P. Wang (2001) Spatial mismatch in search equilibrium, Journal of Labor Economics,

19, 949-72.

[12] Crane J. (1991) The epidemic theory of ghettos and neighborhood effects on dropping out and teenage child-

bearing, American Journal of Sociology, 96, 1226-1259.

[13] Dahl G. (2002) Mobility and the return to education: testing a Roy model with multiple markets, Economet-

rica, 70, 2367-2420.

[14] Davies S. and D. Huff (1972) Impact of ghettoization on black employment, Economic Geography, 48, 421-427.

[15] Dawkins C., Shen Q. and T. Sanchez (2005) Race, Space and Unemployment Duration, Journal of Urban

Economics, 58, 91-113.

21

[16] Desplanques G. (1994) Connaître les migrations, Espace, Populations et Sociétés, 1, 34-39.

[17] DREIF (Direction Régionale de l’Equipement Ile-de-France) / INSEE Ile-de-France (1997), “Déplacements

quotidiens : les Franciliens plus mobiles”, 2 pages.

[18] Dujardin C. and F. Goffette-Nagot (2005) Neighborhood effects, public housing and unemployment in a French

city, document de travail du GATE 05-05, 35 pages.

[19] Dujardin C., Selod H. and I. Thomas (2007) Residential Segregation and Unemployment: the case of Brussels,

Urban Studies, forthcoming.

[20] Duncan O. and B. Duncan (1955) A methodological analysis of segregation indexes, American Sociological

Review, 41, 210-217.

[21] Ferguson R. (1991) Paying for public education: new evidence on how and why money matters, Harvard

Journal on Legislation, 28, 465-98.

[22] Fieldhouse E. (1999) Ethnic minority unemployment and spatial mismatch: the case of London, Urban Studies,

36, 1569-96.

[23] Fitoussi J.P., Laurent E. and J. Maurice (2004), Ségrégation Urbaine et Intégration Sociale, rapport du Conseil

d’Analyse Economique n◦45, La Documentation Française, Paris.

[24] Geurs K. and J. Ritsema van Eck (2001) Accessibility measures: review and applications, RIVM report 408505

006, 265 p.

[25] Gilbert G. (2004) L’économie de la dépense publique locale, in J.F. Thisse, F. Maurel, A. Perrot, J.C. Prager

and J.P. Puig (eds.), Villes et Economie, Paris: La Documentation Française.

[26] Gobillon L. (2001) Emploi, logement et mobilité résidentielle, Economie et Statistique, 349-350, 77-98.

[27] Gobillon L. and H. Selod (2004) Les déterminants spatiaux du chômage en Ile-de-France, in

Ségrégation Urbaine et Intégration Sociale, Fitoussi, Laurent et Maurice, eds., rapport du Conseil d’Analyse

Economique n◦45, La Documentation Française, Paris, 171-187.

[28] Gobillon L., Selod H. and Y. Zenou (2007) The mechanisms of spatial mismatch, Urban Studies, forthcoming

(in Volume 44, Issue 13, December).

[29] Goux D. and E. Maurin (2004) Neighborhood effects on performance at school, mimeo.

[30] Harding D. (2003) Counterfactual models of neighborhood effects: the effect of neighborhood poverty on

dropping out and teenage pregnancy, American Journal of Sociology, 109, 676-719.

[31] Héran, F. (1987) Comment les Français voisinent, Economie et Statistique, 195, 43-59.

22

[32] Holzer H. (1987) Informal job search and black youth unemployment, American Economic Review, 77, 446-452.

[33] Holzer H. (1988) Search method used by unemployed youth, Journal of Labor Economics, 6, 1-20.

[34] Holzer H., Ihlanfeldt K. and D. Sjoquist (1994), “Work, search and travel among white and black youth”,

Journal of Urban Economics, 35, 320-345.

[35] Hoxby C. (2000) The effects of class size on student achievement: new evidence from population variation,

Quarterly Journal of Economics, 115, 1239-1285.

[36] Ihlanfeldt K. (1997) Information on the spatial distribution of job opportunities within Metropolitan Areas,

Journal of Urban Economics, 41, 218-242.

[37] Ihlanfeldt K. and D. Sjoquist (1998) The spatial mismatch hypothesis: a review of recent studies and their

implications for welfare reform, Housing Policy Debate, 9, 849-892.

[38] Imbens G. (2003) Sensitivity to exogeneity assumptions in program evaluation, American Economic Review,

93(2), 126-132.

[39] Jencks C. and S. Mayer (1990) The social consequences of growing up in a poor neighborhood, in L. Lynn and

M. McMgreary (eds.), Inner City Poverty in the United States, Washington DC: National Academy Press.

[40] Joseph A. and P. Bantock (1982) Measuring potential physical accessibility to general practitioners in rural

areas: a method and case study, Social Science Medecine, 16, 85-90.

[41] Kain J. (1992) The spatial mismatch hypothesis: three decades later, Housing Policy Debate, 3, 371-460.

[42] Magnac T., Robin J.-M. and M. Visser (1995) Analysing incomplete individual employment histories using

indirect inference, Journal of Applied Econometrics, 10, 153-169.

[43] Mortensen D. and T. Vishwanath (1994) Personal contacts and earnings. It is who you know!, Labour Eco-

nomics, 1, 187-201.

[44] Ortega J. (2000) Pareto-improving immigration in an economy with equilibrium unemployment, Economic

Journal, 110, 92-112.

[45] Patacchini E. and Y. Zenou (2006) Search activities, cost of living and local labor markets in Britain, Regional

Science and Urban Economics, 36, 227-248.

[46] Phelps E. (1972) The statistical theory of racism and sexism, American Economic Review, 62, 659-61.

[47] Piketty T. (2004) L’Impact de la taille des classes et de la ségrégation sociale sur la réussite scolaire dans les

écoles françaises : une estimation à partir du panel primaire 1997, mimeo, 71 p.

[48] Préteceille E. (2003) La division sociale de l’espace francilien. Typologie socioprofessionnelle 1999 et transfor-

mations de l’espace résidentiel 1990-99, Paris: Observatoire Sociologique du changement, 148 pages.

23

[49] Reingold D. (1999) Social networks and the employment problem of the urban poor, Urban Studies, 36, 1907-

1932.

[50] Rogers C. (1997) Job search and unemployment duration: Implications for the spatial mismatch hypothesis,

Journal of Urban Economics, 42, 109-132.

[51] Rosenbaum P. and D. Rubin (1983) Assessing sensitivity to an unobserved binary covariate in an observational

study with binary outcome, Journal of the Royal Statistical Society B, 45, 212-218.

[52] Schwartz (1973) Interpreting the Effect of Distance on Migration, The Journal of Political Economy, 81,

1153-1169.

[53] Selod H. and Y. Zenou (2001) Social interactions, ethnic minorities and urban unemployment, Annales

d’Economie et de Statistique, 63-64, 183-214.

[54] Selod H. and Y. Zenou (2006) City Structure, Labor Discrimination and Job Search. Theory and Policy

Implications, Economic Journal, 116, 1057-87.

[55] Smith T. and Y. Zenou (2003) Spatial Mismatch, Search Effort, and Urban Spatial Structure, Journal of

Urban Economics, 54, 129-156.

[56] Turner S. (1997) Barriers to a better break: Employer discrimination and spatial mismatch in metropolitan

Detroit, Journal of Urban Affairs, 19, 123-141.

[57] Wasmer E. and Y. Zenou (2002) Does city structure affect search and welfare?, Journal of Urban Economics,

51, 515-541.

[58] Wasmer E. and Y. Zenou (2005) Equilibrium search unemployment with explicit spatial frictions, Labour

Economics, 13, 143-165.

[59] Weinberg B. (2000) Black residential centralization and the spatial mismatch hypothesis, Journal of Urban

Economics, 48, 110-134.

[60] Weinberg B. (2004) Testing the spatial mismatch hypothesis using inter-city variations in industrial composi-

tion, Regional Science and Urban Economics, 34, 505-32.

[61] Wenglenski S. (2003), Une mesure des disparités socio-spatiales d’accessibilité au marché de l’emploi en Ile-

de-France, thèse de l’Université Paris XII-Val-de-Marne-CRETEIL-LVMT-INRETS

[62] Wilson J. (1987) The Truly Disadvantaged: the Inner City, the Underclass, and Public Policy. Chicago:

University of Chicago Press.

[63] Zenou Y. (2002) How do firms redline workers?, Journal of Urban Economics, 52, 391-408.

24

7 Appendix A: Alternative measures of job accessibility

The computation of ratios of local labor demand and supply provides a municipality-wise measure of job-accessibility.

It quantifies the possible mismatch between local labor demand and labor supply. We provide three different in-

dicators depending on the assumptions that can be made to compute the local labor supply and demand. All our

indices are computed using job and worker locations from the 1999 Census of the Population and the commuting

times between municipalities estimated at morning peak hours in 2000 by the Ministry of Transport for both public

and private transport modes.19

Our first index simply measures the ratio of available jobs to resident workers (whether occupied or not) within

a 45-minute commuting trip around each municipality. This is the index we used throughout the present study as

its meaning is straightforward. The formula for the index is as follows:

Ai1 =

NPj=1f (dij)Ej

NPk=1

f (dik)Rk

=NXj=1

f (dij)

⎛⎜⎜⎜⎝ EjNPk=1

f (dik)Rk

⎞⎟⎟⎟⎠ (5)

where i is the considered municipality, N is the number of municipality, Ej is the number of jobs in municipality j,

Rj is the number of resident workers in j, and f (dij) is a distant decay function with dij the travel time between

i and j (where f (dij) is equal to 1 if dij is equal to or lower than 45 minutes, and 0 otherwise). This indicator

is a very simple measure of job accessibility that should be considered only as a first-order approximation of the

local labor-market tension in each municipality. Indeed, even though it provides a correct measure of the relevant

labor demand (by counting all the jobs that can be reached in 45 minutes or less from location i) it provides only

an approximation of the relevant local labor supply by assuming that only the workers who reside within the same

45-minute isochron centered on i are competing for these jobs.

To address this issue, we also use a second indicator that was previously proposed by Joseph and Bentock

(1982) in another context and which incorporates a more appropriate measure of the local labor supply:

Ai2 =NXj=1

f (dij)

⎛⎜⎜⎜⎝ EjNPk=1

f (dkj)Rk

⎞⎟⎟⎟⎠ (6)

The main difference with (5) lies in the second term in (6) which better captures the competition for jobs in a

municipality j accessible in 45 minutes or less from municipality i. Indeed, for each municipality j accessible within

45 minutes of a municipality i, the local supply termNPk=1

f (dkj)Rk is now correctly centered on j and not on i as

in (5). It thus correctly considers all the workers that could travel to work in j within 45 minutes. However, the

correction for the competion factor is not entirely satisfying as it can be seen from (6) that some workers may be

19Matrices of commuting time between municipalities were estimated by the Ministry of Transport (Direction Regionale de

l’Equipement Ile-de-France) using the General Transport Survey (Enquête Globale des Transports, 2000). For a detailed description of

this dataset and its use to calculate other accessibility indices, see Wenglenski (2003).

25

counted several times as they may compete for several jobs within the isochron centered on i (Geurs and Ritsema

van Eck, 2001).

Consequently, we propose a third index which is a generalized and modified version of the Joseph and Bantock

indicator. It introduces an imputed job-search intensity correction that affects both the relevant labor supply and

demand according to the following formula:

Ai3 =NXj=1

pijEj

NPk=1

pkjRk

(7)

with

pij =f (dij)EjNPj=1f (dij)Ej

(8)

where pij is the fraction of time that individual i devotes to search in location j (we haveNPj=1pij = 1). (8)

assumes that workers search more in neighborhoods where jobs are concentrated. For each municipality, this index

corresponds to a ratio of effective labor demand and labor supply assuming that workers search only within a

45-minute isochron of their residence and that they spend proportionally more time searching in places where jobs

are more numerous.

These indices can be calculated for each transport mode (since the 45-minute isochrons differ). For private

transport, they are highly correlated, the correlation between A1 and A2 (resp. A1 and A3, and A2 and A3) being

.86 (resp. .88 and .86). For public transport, the correlations are lower. For A1 and A2 (resp. A1 and A3, and A2

and A3), the correlation is .31 (resp. .21 and .31).

8 Appendix B: the sensitivity analysis

The estimation of parameters is conducted using maximum likelihood. For any given individual i, the probabilities

in this appendix are conditional on the set of observed explanatory variables in the transition and location equations.

To make notations simple, this conditioning is omitted. In the likelihood, the contribution of a given individual

depends on his number of transitions. When an individual only experiences one transition at date t0, his contribution

writes:

Pi = P (Eit0 = e0, Li = j)

where e0 is the observed exit. When an individual experiences two transitions at dates t0 and t0+1, his contribution

writes:

Pi = P (Eit0 = e0, Eit0+1 = e1, Li = j)

where e0 and e1 are the observed exits at dates t0 and t0+1.

26

The log-likelihood writes Λ =PilnPi. It can be decomposed as Λ = ΛL + ΛE|L where ΛL is the log-likelihood of

location:

ΛL =Xi

lnP (Li = j)

and ΛE|L is the log-likelihood of exit conditionally on location:

ΛE|L =Xi∈Φ

lnP (Eit0 = e0 |Li = j ) +Xi∈Ψ

lnP (Eit0 = e0,Eit0+1 = e1, |Li = j )

where Φ is the set of individuals experiencing one transition only and Ψ is the set of individuals experiencing two

transitions.

The log-likelihood of location ΛL is very easy to compute. The individual contribution to the conditional log-

likelihood ΛE|L can be computed by generalizing the calculations in Rosenbaum and Rubin (1983). Calculations

are detailed below in subsection 8.1. As the number of locations is small (J = 8) compared to the number of

observations, it is possible to estimate the coefficients of all observed variables in the location equations. We

maximize the full likelihood Λ with respect to α = (α2, ...,αE), β = (β2, ...,βE), λ = (λ2, ...,λJ) and θ.

8.1 Calculation of the log-likelihood

The calculation of the likelihood depends on the law of the unobserved individual variable. In the location log-

likelihood ΛL, the contribution of individual i writes:

P (Li = j) = Eui [P (Li = j |ui )] (9)

with:

P (Li = j |ui ) =exp(Ait0λj +Bjt0θ + δjui)

1 +X`

exp(Ait0λ` +B`t0θ + δ`ui)(10)

If individual i experiences one transition only, his contribution to the conditional likelihood ΛE|L writes:

P (Eit0 = e0 |Li = j ) = Eui|Li=j [P (Eit0 = e0 |Li = j, ui )] (11)

If individual i experiences two transitions, using the fact that, conditionally on ui, exits are independent, his

contribution to the conditional likelihood ΛE|L writes:

P (Eit0 = e0, Eit0+1 = e1 |Li = j ) = Eui|Li=j [P (Eit0 = e0 |Li = j, ui )P (Eit0+1 = e1 |Li = j, ui )] (12)

where:

P (Eit = e |Li = j, ui ) =exp (Xitαe + Zjtβe + γeui)

1 +Xe>1

exp (Xitαe + Zjtβe + γeui)(13)

Denote f(.) the density of ui and f(. |Li = j ) the density of ui conditionally on location. Using the Bayes rule, wehave:

27

f (ui |Li = j ) =P (Li = j |ui ) f (ui)

P (Li = j)(14)

Plugging (9) and (10) into (14), we get a formula for f (ui |Li = j ) that only depends on observed explanatoryvariables and parameters. Using this formula as well as (13), and (11) or (12), we obtain the contribution of

individual i to the conditional likelihood.

In practice, following the literature (Rosenbaum and Rubin, 1983) we assume that ui simply follows a Bernoulli

law, taking two values, a1 and a2, with respective probabilities π1 and π2. We can then simplify the expressions.

In the location likelihood ΛL, the contribution of individual i writes:

P (Li = j) =2X

k=1

πkP (Li = j |ui = ak ) (15)

with:

P (Li = j |ui = ak ) =exp(Ait0λj +Bjt0θ + δjak)

1 +X`

exp(Ait0λ` +B`t0θ + δ`ak)(16)

If individual i experiences one transition only, his contribution to the conditional likelihood ΛE|L writes:

P (Eit0 = e0 |Li = j ) =2X

k=1

P (Eit0 = e0 |Li = j, ui = ak )P (ui = ak |Li = j ) (17)

If individual i experiences two transitions, his contribution to the conditional likelihood ΛE|L writes:

P (Eit0 = e0, Eit0+1 = e1 |Li = j )

=2Xk=1

P (Eit0 = e0 |Li = j, ui = ak )P (Eit0+1 = e1 |Li = j, ui = ak )P (ui = ak |Li = j ) (18)

The probability that individual i experiences an exit of type e conditionally on the unobserved individual variable

is given by:

P (Eit = e |Li = j, ui = ak ) =exp (Xitαe + Zjtβe + γeak)

1 +Xe>1

exp (Xitαe + Zjtβe + γeak)(19)

Using the Bayes rule, we get for any k:

P (ui = ak |Li = j ) =P (ui = ak)P (Li = j |ui = ak )

P (Li = j)(20)

Using equations (20), (15), (19), and (17) or (18), we obtain the contribution of individual i to the conditional

likelihood.

8.2 Application

We apply the methodology to the specification of labor-market transitions that uses seven dummies for the HAC

clusters (J = 8).

28

We assume that ui is a binary variable that is drawn in the set {−1/2, 1/2} with a probability equal to 1/2 for eachvalue. In Ai, we include a constant and dummies for gender, age and age2, diploma, and nationality. We do not

include any local variable in Bj since the values would not vary much across individuals (as there are only eight

clusters).

We test the robustness of the results on four sets of S = 1, 000 simulations for the effects of individual unobserved

heterogeneity. In the first set, the coefficients of the unobserved variable (γe and δj) are drawn in the uniform

law on [− ln 2; ln 2] ensuring that the corresponding odds-ratios are comprised between 1/2 and 2. In the second,third, and four sets of simulations, they are drawn in the uniform laws on [− ln 3; ln 3], [− ln 4; ln 4] and [− ln 5; ln 5]respectively. The coefficients for the transition and location of reference are always set at 0 (δ0 = 0 and γ0 = 0).

Simulation results for odds-ratios in [−1/2, 1/2] (γe and δj drawn in [− ln 2; ln 2]) are reported in Tables A2to A4.Graph 1 represents the descriptive statistics on the sign and significance of the coefficient of the type-III

cluster for each labor-market transition. Graph 2 plots the same information for the type-V cluster. Both graphs

are discussed in the main body of the text.

[Insert Tables A2 to A4]

9 Appendix C: Alternative specifications

We ran alternative specifications of equation (2).20

First, we discretized the segregation and job-accessibility variables in order to detect potential non-linearities.

For each contextual variable, we created dummies for quartiles and re-ran the regression with those dummies. The

are three noticeable differences with the main regression. The municipality proportion of Africans is not significant

anymore. There is a non-linearity effect for the job-density by public transport: the third quartile has a significant

negative effect on finding a job (the reference being the first quartile) but the fourth quartile has no effect. For the

job density by private transport, the fourth quartile has a significant positive effect on the premature exit from the

panel.

Second, we changed the geographic scale of the segregation variables:

• When the proportion of educated workers in the neighborhood is replaced by the same proportion but at themunicipality level, none of the segregation variables has a significant effect on finding a job and prematurely

exiting the panel. Both variables have a negative significant effect on dropping out of the labor force. These

results could be explained by the negative correlation between the proportion of Africans and educated

workers at the municipality level.

• Conversely, when the proportion of Africans is measured at the neighborhood level rather than at the mu-nicipality level, we obtain the same result for finding a job except that the negative effect of job density by

20Details are available upon request.

29

public transport is now significant at a 5% level (instead of 10%). For the premature exit from the panel,

the proportion of Africans still has a negative effect but it becomes non-significant. For dropping out of the

labor force, job density by public transport has a negative effect that is now significant at 10%.

• When the proportion of educated workers is at the municipality level and the proportion of Africans is at theneighborhood level, for finding a job, the signs of the contextual effects remain the same. However, the effect

of the proportion of educated workers is not significant anymore. For prematurely exiting from the panel, the

signs are also unchanged but the proportion of Africans does not have a significant effect anymore.

Third, we also considered other segregation variables:

• When the neighborhood proportion of educated workers is replaced by the neighborhood proportion of high-skilled white collars, we find very similar results. This is not surprising since the two variables are highly

correlated. For finding a job, the neighborhood proportion of high-skilled white collars has a positive effect

that is significant at 10%. The job density by public transport still has a negative effect that is significant at

10%. For the premature exit from the panel, the neighborhood proportion of high-skilled white collars still

has a positive effect but it is not significant. The municipality proportion of Africans has a negative effect

which becomes significant at 5%.

• When the neighborhood proportion of educated workers is replaced by the neighborhood proportion of unem-ployed and out-of-the-labor-force residents, the latter has a negative effect on finding a job that is significant

at 1%.21 This result is consistent with Dujardin, Selod and Thomas (2004) on Brussels who show that the

local unemployment rate has a positive effect the unemployment probability of young adults living with their

parents. Job density by public transport does not have a significant effect anymore. For the premature exit

from the panel, the municipality proportion of Africans has a negative effect which becomes significant at 5%.

The job density by private transport has a positive effect which becomes significant at 10%. The proportion

of unemployed and out-of-the-labor-force neighbors has a negative effect but it is not significant even at the

10% level.

Fourth, we used the alternative job-accessibility indices described in appendix A. When we use the measure

A2, results remain the same for segregation variables. The negative effect of job accessibility by private transport

on finding a job becomes significant at the 10% level. When we use the measure A3, the municipality proportion

of Africans still has negative effects on all exits which become significant at the 5% level except for dropping out

of the labour which is significant at 10%. Once again, the negative effect of job accessibility by private transport

on finding a job becomes significant at the 10% level.

21One should be cautious when interpreting this result. Indeed, the effect can be due to the weakness of social interactions or to

redlining. It can also be generated by some local unobservables that would simultaneously explain the local unemployment rate and

the local constraints impeding unemployed workers from finding a job.

30

Fifth, we also computed job densities separately by educational level. In the regression, we crossed a dummy

for each educational level (with vs without a high school diploma) with the corresponding job-density and with

the segregation variables. For finding a job, we find that only the low-educated are positively affected by the

neighborhood level of education. The low-educated are the only ones to be negatively affected (at the 10% level)

by their job accessibility.

31

Map 1. Municipality unemployment rates in the Paris region

< 5% 5 – 10% 10 – 15% 15 – 20 % 20 – 25%

Source: Census of the Population (1999), INSEE. Note: Geographical units are municipalities. The thick lines are the borders of the region’s districts (départements). The city of Paris is at the center. Clockwise around Paris are the Hauts-de-Seine (west), the Seine-Saint-Denis (north east) and the Val-de-Marne (south east). The peripheral districts are the Yvelines (west), the Val d'Oise (north), the Seine-et-Marne (east), and the Essonne (south).

Map 2. Specialization index for educated workers

(at the municipality level)

< .5 .5 – .8 .8 – 1.2 1.2 – 1.5 1.5 – 2.1

Source: Census of the Population (1999), INSEE. Note: Educated workers have a high school diploma or above.

Table 1. Residential segregation indexes Paris

region City of Paris

Inner ring Outer ring

Inter-municipality segregation indexes Unemployed / occupied workers .16 .12 .17 .15 Skilled white collars / clerks .26 .14 .26 .21 Skilled white collars / blue collars .39 .23 .38 .32 Blue collars / clerks .15 .10 .14 .13 Educated workers (high school diploma) / Low-educated workers

.25 .15 .23 .19

French / Maghrebines .32 .21 .27 .39 French / Africans (excl. Maghrebines) .32 .23 .25 .39 Africans (excl. Maghrebines) / Maghrebines

.13 .05 .12 .19

Source: Census of the Population (1999), INSEE. Note: The indexes are Duncan and Duncan (1955) computed on a spatial breakdown with municipalities as geographical units.

Map 3. Residential specialization index for Africans (at the municipality level)

< .5 .5 – .8 .8 – 1.2 1.2 – 1.5 1.5 – 3.4

Source: Census of the Population (1999), INSEE.

Table 2. Job disconnection and accessibility measures Paris

regionCity of Paris

Inner ring

Outer ring

Inter-municipality job-disconnection indexes Labor force / jobs .25 .22 .20 .25 Skilled white collars / corresponding jobs .31 .23 .30 .39 Clerks / corresponding jobs .30 .27 .21 .26 Blue collars / corresponding jobs .25 .24 .18 .25 Educated workers (high-school diploma or above) /

corresponding jobs .26 .21 .24 .30

Low-educated workers (without high school diploma) / corresponding jobs

.26 .26 .19 .23

French / jobs held by French workers .26 .22 .22 .27 Maghrebines / jobs held by Maghrebine workers .25 .25 .17 .27 Africans (excl. Maghrebines) / jobs held by African (excl. Maghrebines) workers

.33 .32 .25 .30

Average job-density within 45 minutes By public transport

All 1.06 1.30 1.22 .81 High-school diploma 1.07 1.21 1.20 .85 Without high-school diploma 1.05 1.44 1.26 .77

By private transport All .86 1.03 .87 .76 High-school diploma .91 1.05 .91 .81 Without high-school diploma .80 1.00 .83 .72

Sources: Census of the Population (1999), INSEE, and General Transport Survey (2000), DREIF. Notes: Duncan and Duncan (1955) indexes are computed on a municipality spatial breakdown. Averages are calculated weighting by the labor force in each municipality.

Map 4. Job densities within 45 minutes by public transport (centered on each municipality)

< .5 .5 – .8 .8 – 1.22 1.2 – 2.0 2.0 – 19.9

Sources: Census of the Population (1999), INSEE, and General Transport Survey (2000), DREIF.

Map 5. Job densities within 45 minutes by private transport

(centered on each municipality)

< .5 .5 – .8 .8 – 1.2

Sources: Census of the Population (1999), INSEE, and General Transport Survey (2000), DREIF.

Map 6. Number of observed labor-market transitions in the 1990-2002 Labor Force Survey

< 15 15 – 29 30 – 74 75 – 124 125 - 271

Source: Labor Force Survey (1990-2002), INSEE Note: A municipality may host several neighborhoods.

Table 3. Descriptive statistics on neighborhoods-years (1990-2002 LFS) Average Std dev Min Max 1st

decile 9th

decile Totals

Number of people 27.0 15.1 1 109 12 48 Number of workers in the labor force

21.8 12.2 1 89 10 38

Number of unemployed or out-of-the-labor-force individuals

7.8 5.2 1 49 2 15

Number of unemployed workers 2.5 1.9 1 18 1 5 Percentages

% unemployed or out-of-the-labor-force individuals

.29 .13 .04 1.00 .14 .46

% unemployed workers .13 .10 .01 1.00 .05 .25 % skilled white collars .21 .20 .00 1.00 .00 .50 % educated (high school diploma) .41 .24 .00 1.00 .13 .75 % Africans .07 .12 .00 1.00 .00 .23

Source: Labor Force Survey (1990-2002), INSEE. Note: There are 3,905 observations of the neighborhood-year type for which labor-market transitions starting during the 1990-2001 period are observed. Averages are weighted by the number of transitions.

Table 4. Correlation matrix between neighborhood and municipality variables Neighborhoods Municipalities

% skilled white

collars

% educa

ted worke

rs

% Africans

(including Maghrebines

)

% out-of-the-labor-force

or unemploy

ed workers

% skilled white

collars

% educated workes

%

Africans (includin

g Maghrebines)

%

unemployed workers

Job density

within 45 minutes (private

vehicules)

Job density

within 45 minutes (public

transport) Neighborhoods

% skilled white collar neighbors 1.00

%educated neighbors .82 1.00

% African neighbors -.35 -.34 1.00

% out-of-the-labor-force or unemployed neighbors -.28 -.30 .46 1.00

Municipalities % skilled white collars in the municipality .60 .63 -.22 -.22 1.00

% educated workers in the municipality .60 .63 -.21 -.27 .98

1.00

% Africans in the municipality -.32 -.27 .48 .23 -.47

-.46

1.00

% unemployed workers in the municipality -.36 -.33 .42 .36 -.58

-.56

.90

1.00

Job density within 45 minutes (private transport) .39 .41 .07 -.02 .60

.60

. 17

.03 1.00

Job density within 45 minutes (public transport) .20 .23 .13 .05 .32

.32

.32

.23 .64 1.00

Sources: Labor Force Survey (1999-2002) and Census of the Population (1999), INSEE, and General Transport Survey (2000), DREIF. Note: Statistics are computed by weighting by the number of transitions in each neighborhood-area. All correlations are significant at the 1% level.

Table 5a. The Hierachical Ascending Classifications of neighborhoods Neighborhood

segregation Municipality segregation

Job accessibility

strong weak strong weak bad good Neighborhood characteristics

% educated neighbors

26.4 63.3

% African neighbors

11.7 1.3

Municipality characteristics

% educated in municipality

34.5 50.8

% Africans in municipality

12.8 4.5

Job-accessibility indexes

Job density within 45 minutes (public)

.88 1.38

Job density within 45 minutes (private)

.77 1.02

Totals Number of areas

1,457 1,639 1,402 1,694 1,660 1,436

Number of areas with transitions

1,264 1,160 1,157 1,267 1,346 1,078

Number of transitions

6,520 3,120 5,448 4,192 5,074 3,566

Sources: Labor Force Survey (1999-2002), Census of the Population (1999), INSEE, and General Transport Survey (2000), DREIF.

Table 5b. The eight-class typology of neighborhoods I II III IV V VI VII VIII Class characteristics

Neighborhood segregation weak weak strong strong weak weak strong strongMunicipality segregation weak strong weak strong weak strong weak strongJob accessibility good good good good bad bad bad bad

Neighborhood characteristics

% educated neighbors 70.8 64.9 34.6 29.9 56.2 52.7 23.7 23.7 % African neighbors 1.7 1.8 12.3 18.2 .4 1.0 5.3 12.3

Municipality characteristics

% educated in municipality

60.6 45.3 55.8 39.7 44.6 32.5 36.1 28.5

% Africans in municipality 5.3 12.8 5.9 15.0 3.5 10.5 3.4 12.6 Job-accessibility indexes

Job accessibility (public) 1.39 1.35 1.44 1.36 .83 .92 .74 .98 Job accessibility (private) 1.03 1.00 1.04 .99 .79 .76 .74 .76

Totals Number of areas 704 253 196 283 442 240 352 626 Number of areas with transitions

490 182 159 247 311 177 307 551

Number of transitions 1,266 501 547 1,252 829 524 1,550 3,171 Sources: Labor Force Survey (1999-2002), Census of the Population (1999), INSEE, and General Transport Surve (2000), DREIF. yNote: “Weak segregation” corresponds to a high proportion of educated workers and a low proportion of Africans.

Map 7. Main neighborhood type in the municipality

-Neighborhood segregation -Municipality segregation -Job accessibility I. weak, weak, good II. weak, strong, good III. strong, weak, good IV. strong, strong, good V. weak, weak, bad VI. weak, strong, bad VII. strong, weak, bad VIII. strong, strong, bad

Sources: Labor Force Survey (1999-2002), Census of the Population (1999), INSEE, and General Transport Survey (2000), DREIF. Note: The baseline neighborhoods (type I) are characterized by weak levels of segregation both at the neighborhood and municipality scales as well as a good job accessibility.

Table 6a. Results of the regression with segregation and job-accessibility measures

Finding a job without moving Dropping out of the labor force without moving

Leaving the panel (move or no answer)

Coeff. (Std dev)

Odds-ratio

Coeff. (Std dev)

Odds-ratio

Coeff. (Std dev)

Odds-ratio

Age/10 .394** (.189)

1.483

-1.830*** (.208)

.160

-.756*** (.185)

.470

(Age/10) squared

-.096*** (.025)

.909

.253*** (.026)

1.288

.046* (.024)

1.047

Gender: Female -.097* (.055)

.908

.699*** (.069)

2.011

-.015 (.055)

.985

Education No diploma or professional certificate (CEP)

<Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.>

Technical diploma (below high school diploma)

.200*** (.071)

1.221

-.065 (.089)

.937

.058 (.073)

1.059

Technical high school diploma

.648*** (.145)

1.912

.023 (.202)

1.024

.274* (.156)

1.316

General high school diploma -.313*** (.119)

.731

-.084 (.151)

.920

-.075 (.115)

.927

University degree .441*** (.082)

1.554

.121 (.107)

1.129

.350*** (.083)

1.419

Nationality French <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> European (other) .227**

(.111) 1.255

.029

(.141) 1.029

.155

(.114) 1.168

Maghrebine -.175*

(.100) .839

.067

(.115) 1.069

.002

(.095) 1.002

African (other) -.258*

(.147) .773

.145

(.174) 1.156

.104

(.134) 1.109

Other nationality -.228

(.141) .796

-.057 (.172)

.945

-.068 (.134)

.934

Search method National Employment Agency (ANPE) only

<Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.>

National Employment Agency (ANPE) and other methods

.313*** (.111)

1.367

-.693*** (.103)

.500

-.130 (.101)

.878

Without National Employment Agency (ANPE)

.637*** (.136)

1.892

.199 (.133)

1.221

.241* (.127)

1.272

Housing occupancy status Owner <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> Renter of the public sector (HLM)

-.041 (.076)

.959

.165* (.093)

1.180

.398*** (.081)

1.489

Renter of the private sector or free occupation

.002 (.071)

1.002

-.053 (.091)

.948

.715*** (.074)

2.044

% neighbors with high school diploma

.356** (.162)

1.427

-.268 (.202)

.765

.454*** (.160)

1.575

% Africans in the municipality

-.690 (.522)

.501

-.827 (.646)

.438

-.936* (.524)

.392

Job density within 45 minutes (public transport)

-.169* (.097)

.845

-.161 (.121)

.851

-.073 (.101)

.929

Job density within 45 minutes (private transport)

-.254 (.259)

.776

-.034 (.317)

.967

.283 (.263)

1.327

Sources: Labor Force Survey (1999-2002), Census of the Population (1999), INSEE, and General Transport Survey (2000), DREIF. Note: Number of observations: 9,640. *** significance at the 1% level; ** 5%, * 10% . We included a constant term and year fixed effects. The coefficients are not reported in the table.

Table 6b. Results of the regression with dummies for neighborhood type

Finding a job without moving Dropping out of the labor force without moving

Leaving the panel (move or no answer)

Coeff. (Std dev)

Odds-ratio

Coeff. (Std dev)

Odds-ratio

Coeff. (Std dev)

Odds-ratio

Age/10 .399** (.189)

1.490

-1.829*** (.208)

.161

-.737*** (.185)

.479

(Age/10) squared

-.096*** (.025)

.908

.253*** (.026)

1.288

.043* (.024)

1.044

Gender: Female -.099* (.055)

.906

.691*** (.069)

1.996

-.014 (.055)

.986

Education No diploma or professional certificate (CEP)

<Ref.> <Ref.>

<Ref.> <Ref.>

<Ref.> <Ref.>

Technical diploma (below high school diploma)

.207*** (.071)

1.230

-.067 (.089)

.935

.062 (.073)

1.064

Technical high school diploma

.615*** (.146)

1.850

.008 (.202)

1.008

.237 (.157)

1.267

General high school diploma -.333*** (.119)

.717

-.093 (.151)

.911

-.101 (.115)

.904

University degree .454*** (.083)

1.575

.111 (.107)

1.117

.378*** (.083)

1.459

Nationality French <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> European (other) .219**

(.111) 1.245

.032

(.141) 1.033

.154

(.114) 1.166

Maghrebine -.195*

(.100) .823

.062

(.115) 1.064

-.009 (.095)

.991

African (other) -.276* (.148)

.759

.129 (.174)

1.138

.091 (.134)

1.095

Other nationality -.240 (.141)

.787

-.058 (.172)

.944

-.077 (.134)

.926

Search method National Employment Agency (ANPE) only

<Ref.> <Ref.>

<Ref.> <Ref.> <Ref.> <Ref.>

National Employment Agency (ANPE) and other methods

.303*** (.111)

1.354

-.695*** (.103)

.499

-.144 (.101)

.866

Without National Employment Agency (ANPE)

.638*** (.136)

1.893

.210 (.133)

1.234

.233* (.127)

1.262

Housing occupancy status Owner <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> Renter of the public sector (HLM)

-.041 (.075)

.960

.221** (.091)

1.247

.414*** (.081)

1.513

Renter of the private sector or free occupation

.020 (.072)

1.020

-.023 (.091)

.977

.742*** (.075)

2.100

Neighborhood type Type I <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> Type II -.231

(.143) .794

-.091 (.186)

.913

-.158 (.137)

.854

Type III -.341** (.149)

.711

.011 (.178)

1.011

-.103 (.135)

.902

Type IV -.042 (.115)

.959

.062 (.148)

1.064

-.098 (.111)

.907

Type V .338*** (.124)

1.402

.394** (.157)

1.483

.234* (.126)

1.264

Type VI .277** (.141)

1.319

.333 (.183)

1.395

.169 (.145)

1.184

Type VII .030 (.110)

1.030

.285** (.139)

1.330

-.185* (.111)

.831

Type VIII -.086 (.102)

.918

.140 (.131)

1.150

-.309*** (.101)

.734

Sources: Labor Force Survey (1999-2002), Census of the Population (1999), INSEE, and General Transport Survey (2000), DREIF. Note: Number of observations: 9,640. *** significance at the 1% level; ** 5%, * 10% . We included a constant term and year fixed effects. The coefficients are not reported in the table.

Table 7a. Results of the regression for the sub-sample of workers residing in public housing, with segregation and job-accessibility measures

Finding a job without moving Dropping out of the labor force without moving

Leaving the panel (move or no answer)

Coeff. (Std dev)

Odds-ratio

Coeff. (Std dev)

Coeff. (Std dev)

Odds-ratio

Coeff. (Std dev)

Age/10 .422 (.331)

1.525

-1.705*** (.337)

.182

-.499 (.323)

.607

(Age/10) squared

-.095** (.045)

.909

.226*** (.043)

1.254

.010 (.043)

1.010

Gender: Female .001 (.097)

1.001

.718*** (.116)

2.051

.080 (.098)

1.083

Education No diploma or professional certificate (CEP)

<Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.>

Technical diploma (below high school diploma)

.133 (.118)

1.142

-.115 (.144)

.891

.076 (.121)

1.079

Technical high school diploma

.513 (.348)

1.670

.202 (.432)

1.224

.681** (.334)

1.976

General high school diploma -.330 (.274)

.719

-.030 (.325)

.970

-.329 (.267)

.719

University degree .498** (.196)

1.646

.245 (.241)

1.278

.634*** (.194)

1.885

Nationality French <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> European (other) .007

(.245) 1.007

.163

(.271) 1.177

.279

(.237) 1.322

Maghrebine -.119

(.145) .888

.192

(.157) 1.212

-.036 (.147)

.965

African (other) -.212 (.212)

.809

-.242 (.263)

.785

-.407* (.231)

.666

Other nationality -.088 (.271)

.916

.063 (.295)

1.065

-.070 (.273)

.932

Search method National Employment Agency (ANPE) only

<Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.>

National Employment Agency (ANPE) and other methods

.599*** (.191)

1.820

-.565*** (.164)

.569

-.271* (.156)

.763

Without National Employment Agency (ANPE)

.871*** (.250)

2.390

.250 (.230)

1.283

-.113 (.229)

.893

% neighbors with high school diploma

.777** (.377)

2.175

.115 (.444)

1.122

.837** (.374)

2.308

% Africans in the municipality

-.505 (.897)

.604

-.660 (1.053)

.517

-2.221** (.907)

.108

Job density within 45 minutes (public transport)

-.260 (.199)

.771

.071 (.234)

1.073

-.250 (.202)

.779

Job density within 45 minutes (private transport)

.330 (.496)

1.390

-.110 (.571)

.896

1.187** (.493)

3.276

Sources: Labor Force Survey (1999-2002), Census of the Population (1999), INSEE, and General Transport Survey (2000), DREIF. Note: Number of observations: 3,108. *** significance at the 1% level; ** 5%, * 10% . We included a constant term and year fixed effects. The coefficients are not reported in the table.

Table 7b. Results of the regression for the sub-sample of workers residing in public housing, with dummies for neighborhood type

Finding a job without moving Dropping out of the labor force without moving

Leaving the panel (move or no answer)

Coeff. (Std dev)

Odds-ratio

Coeff. (Std dev)

Coeff. (Std dev)

Odds-ratio

Coeff. (Std dev)

Age/10 .442 (.331)

1.556

-1.694*** (.336)

.184

-.445 (.323)

.641

(Age/10) squared

-.098** (.045)

.907

.226*** (.043)

1.254

.004 (.044)

1.004

Gender: Female -.004 (.097)

.996

.718*** (.116)

2.050

.088 (.098)

1.092

Education No diploma or professional certificate (CEP)

<Ref.> <Ref.>

<Ref.> <Ref.>

<Ref.> <Ref.>

Technical diploma (below high school diploma)

.135 (.118)

1.145

-.119 (.144)

.888

.092 (.121)

1.096

Technical high school diploma

.471 (.349)

1.602

.199 (.433)

1.220

.629* (.334)

1.876

General high school diploma -.343 (.276)

.710

-.026 (.327)

.974

-.327 (.269)

.721

University degree .558*** (.198)

1.747

.246 (.245)

1.279

.679*** (.196)

1.972

Nationality French <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> European (other) -.001

(.245) .999

.139

(.271) 1.149

.274

(.237) 1.315

Maghrebine -.226

(.145) .798

.191

(.157) 1.210

-.041 (.146)

.960

African (other) -.217 (.211)

.805

-.236 (.262)

.790

-.423* (.231)

.655

Other nationality -.084 (.271)

.919

.047 (.295)

1.048

-.075 (.272)

.928

Search method National Employment Agency (ANPE) only

<Ref.> <Ref.>

<Ref.> <Ref.>

<Ref.> <Ref.>

National Employment Agency (ANPE) and other methods

.599*** (.191)

1.820

-.573*** (.164)

.564

-.276* (.156)

.759

Without National Employment Agency (ANPE)

.885*** (.250)

2.423

.251 (.230)

1.285

-.097 (.228)

.908

Neighborhood type Type I <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> Type II -.702

(.559) .496

-.498 (.789)

.608

-.816 (.611)

.442

Type III -.436 (.418)

.647

-.462 (.474)

.630

-.477 (.379)

.621

Type IV -.268 (.386)

.765

-.346 (.439)

.708

-.786** (.357)

.456

Type V .361 (.588)

1.435

-.091 (.713)

.913

.119 (.551)

1.126

Type VI .020 (.538)

1.020

.066 (.602)

1.068

-.580 (.526)

.560

Type VII -.173 (.382)

.841

-.191 (.436)

.826

-.641* (.353)

.527

Type VIII -.314 (.372)

.731

-.278 (.424)

.757

-.877** (.342)

.416

Sources: Labor Force Survey (1999-2002), Census of the Population (1999), INSEE, and General Transport Survey (2000), DREIF. Note: Number of observations: 3,108. *** significance at the 1% level; ** 5%, * 10% . We included a constant term and year fixed effects. The coefficients are not reported in the table.

Table A1. Results of the location model

Neighborhood type II III IV V VI VII VIII Constant -1.478*

(.788) 1.734** (.681)

1.777*** (.569)

1.317** (.599)

-.569 (.729)

1.975*** (.535)

3.008*** (.495)

Age .368 (.411)

-.968*** (.364)

-.358 (.309)

-.705** (.326)

.028 (.397)

-.313 (.291)

-.395 (.268)

Age² -.049 (.051)

.100** (.046)

.006 (.039)

.081** (.041)

-.015 (.051)

.006 (.037)

-.003 (.034)

Female .019 (.117)

-.088 (.116)

.052 (.093)

.139 (.099)

.202* (.117)

.155* (.089)

.114 (.080)

Technical diploma (below high school diploma)

.382** (.188)

.204 (.166)

-.239* (.142)

.249 (.157)

.453** (.175)

.214 (.131)

-.083 (.125)

Technical high school diploma

-.211 (.344)

-.921*** (.350)

-1.222*** (.279)

.442* (.232)

.364 (.267)

-.665*** (.228)

-1.118*** (.222)

General high school diploma

-.180 (.195)

.335 (.227)

.494*** (.178)

.127 (.165)

.682*** (.191)

.542*** (.192)

.879*** (.147)

University degree -.282** (.144)

-1.643*** (.156)

-2.137*** (.126)

-.785*** (.125)

-1.112*** (.156)

-2.558*** (.131)

-2.673*** (.110)

European (other) -.068 (.230)

-.465* (.258)

-.134 (.181)

-.462** (.207)

-.153 (.235)

-.570*** (.180)

-.028 (.152)

Maghrebine .585* (.328)

1.902*** (.255)

2.212*** (.228)

-.609 (.386)

.298 (.345)

.888*** (.242)

1.822*** (.221)

African (other) -.017 (.531)

2.024*** (.333)

2.293*** (.291)

-.516 (.528)

1.018*** (.392)

.985*** (.316)

1.899*** (.277)

Other nationality .271 (.249)

.428* (.258)

.398* (.209)

-.965*** (.314)

.334 (.251)

-.891*** (.264)

.272 (.179)

Sources: Labor Force Survey (1999-2002), Census of the Population (1999), INSEE, and General Transport Survey (2000), DREIF. Note: Number of observations: 9,640. *** significance at the 1% level; ** 5%, * 10% .

Table A2. Results of the sensitivity analysis for the modality “finding a job”, with

dummies for neighborhood type %

<0

% <0 ns

% >0 ns

% >0

Min 1st dec. Med. 9nt dec. Max

Age/10 .000 .000 .000 1.000 .391 .401 .408 .418 .434 (Age/10) Squared 1.000 .000 .000 .000 -.101 -.099 -.098 -.097 -.096 Gender: Female .000 1.000 .000 .000 -.106 -.103 -.100 -.098 -.096 Education No diploma or professional certificate (CEP)

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

Technical diploma (below high school diploma) .000 .000 .000 1.000 .203 .207 .209 .212 .216 Technical high school diploma .000 .000 .000 1.000 .607 .616 .620 .626 .637 General high school diploma 1.000 .000 .000 .000 -.348 -.340 -.336 -.332 -.327 University degree .000 .000 .000 1.000 .439 .452 .457 .462 .476 Nationality French <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> European (other) .000 .000 .929 .071 .210 .216 .219 .222 .227 Maghrebinee .445 .555 .000 .000 -.217 -.202 -.196 -.191 -.176 African (other) .013 .987 .000 .000 -.302 -.285 -.278 -.273 -.255 Other nationality .000 1.000 .000 .000 -.254 -.245 -.241 -.238 -.231 Search method National Employment Agency (ANPE) only

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

National Employment Agency (ANPE) and other methods .000 .000 .000 1.000 .301 .304 .307 .311 .318 Without National Employment Agency (ANPE) .000 .000 .000 1.000 .636 .638 .640 .644 .649 Housing occupancy status Owner <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> Renter of the public sector (HLM) .000 1.000 .000 .000 -.049 -.045 -.042 -.040 -.038 Renter of the private sector or free occupation .000 .000 1.000 .000 .004 .013 .017 .021 .024 Neighbordhood type Type I (reference) <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> Type II .133 .867 .000 .000 -.370 -.287 -.230 -.175 -.100 Type III .858 .142 .000 .000 -.520 -.410 -.339 -.287 -.171 Type IV .000 .848 .152 .000 -.218 -.102 -.041 .012 .096 Type V .000 .000 .020 .980 .184 .288 .341 .393 .483 Type VI .000 .000 .450 .550 .089 .220 .279 .333 .416 Type VII .000 .216 .784 .000 -.118 -.025 .033 .086 .185 Type VIII .001 .989 .010 .000 -.210 -.133 -.086 -.043 .020 Sources: Labor Force Survey (1999-2002), Census of the Population (1999), INSEE, and General Transport Survey (2000), DREIF. Note: Statistics are computed on 1.000 simulations. The first four columns respectively provide the percentage of negative coefficients significant at the 5% level, negative not significant at 5%, positive not significant at 5%, and positive significant at 5%.

Table A3. Results of the sensitivity analysis for the modality “dropping out of the labor force”, with dummies for neighborhood type

%

<0

% <0 ns

% >0 ns

% >0

Min 1st dec. Med. 9nt dec. Max

Age/10 1.000 .000 .000 .000 -1.861 -1.847 -1.838 -1.832 -1.826 (Age/10) Squared .000 .000 .000 1.000 .253 .254 .255 .256 .258 Gender: Female .000 .000 .000 1.000 .691 .694 .696 .700 .705 Education No diploma or professional certificate (CEP)

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

Technical diploma (below high school diploma) .000 1.000 .000 .000 -.076 -.071 -.068 -.065 -.059 Technical high school diploma .000 .994 .006 .000 -.027 -.017 -.011 -.005 .010 General high school diploma .000 1.000 .000 .000 -.102 -.097 -.093 -.089 -.082 University degree .000 .000 1.000 .000 .093 .106 .111 .116 .131 Nationality French <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> European (other) .000 .000 1.000 .000 .021 .027 .031 .034 .040 Maghrebinee .000 .000 1.000 .000 .042 .057 .062 .068 .083 African (other) .000 .000 1.000 .000 .109 .122 .129 .136 .150 Other nationality .000 1.000 .000 .000 -.085 -.073 -.068 -.064 -.056 Search method National Employment Agency (ANPE) only

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

National Employment Agency (ANPE) and other methods 1.000 .000 .000 .000 -.710 -.704 -.699 -.697 -.694 Without National Employment Agency (ANPE) .000 .000 1.000 .000 .202 .206 .209 .211 .213 Housing occupancy status Owner <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> Renter of the public sector (HLM) .000 .000 .000 1.000 .218 .220 .222 .225 .229 Renter of the private sector or free occupation .000 1.000 .000 .000 -.031 -.027 -.024 -.021 -.016 Neighbordhood type Type I (reference) <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> Type II .000 .975 .025 .000 -.276 -.145 -.088 -.032 .078 Type III .000 .347 .653 .000 -.151 -.046 .013 .071 .176 Type IV .000 .101 .899 .000 -.122 .000 .064 .120 .238 Type V .000 .000 .053 .947 .220 .337 .397 .445 .546 Type VI .000 .000 .664 .336 .165 .275 .340 .392 .492 Type VII .000 .000 .343 .657 .127 .226 .287 .345 .416 Type VIII .000 .000 .999 .001 .011 .096 .142 .188 .269 Sources: Labor Force Survey (1999-2002), Census of the Population (1999), INSEE, and General Transport Survey (2000), DREIF. Note: Statistics are computed on 1.000 simulations. The first four columns respectively provide the percentage of negative coefficients significant at the 5% level, negative not significant at 5%, positive not significant at 5%, and positive significant at 5%.

Table A4. Results of the sensitivity analysis for the modality “prematurely leaving the panel”, with dummies for neighborhood type %

<0

% <0 ns

% >0 ns

% >0

Min 1st dec. Med. 9nt dec. Max

Age/10 1.000 .000 .000 .000 -.763 -.750 -.741 -.733 -.725 (Age/10) squared .000 .000 1.000 .000 .041 .043 .044 .045 .047 Gender: Female .000 1.000 .000 .000 -.023 -.017 -.015 -.013 -.010 Education No diploma or professional certificate (CEP)

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

Technical diploma (below high school diploma) .000 .000 1.000 .000 .053 .059 .062 .064 .067 Technical high school diploma .000 .000 1.000 .000 .217 .232 .237 .242 .251 General high school diploma .000 1.000 .000 .000 -.111 -.104 -.100 -.097 -.090 University degree .000 .000 .000 1.000 .354 .375 .380 .385 .401 Nationality French <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> European (other) .000 .000 1.000 .000 .148 .152 .155 .158 .163 Maghrebine .000 .966 .034 .000 -.029 -.014 -.010 -.004 .012 African (other) .000 .000 1.000 .000 .075 .086 .092 .098 .116 Other nationality .000 1.000 .000 .000 -.087 -.079 -.076 -.072 -.065 Search method National Employment Agency (ANPE) only

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

<Ref.>

National Employment Agency (ANPE) and other methods .000 1.000 .000 .000 -.153 -.148 -.145 -.142 -.134 Without National Employment Agency (ANPE) .000 .000 1.000 .000 .225 .230 .232 .234 .239 Housing occupancy status Owner <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> Renter of the public sector (HLM) .000 .000 .000 1.000 .412 .414 .416 .419 .425 Renter of the private sector or free occupation .000 .000 .000 1.000 .743 .745 .748 .752 .757 Neighbordhood type Type I (reference) <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> <Ref.> Type II .014 .985 .001 .000 -.319 -.213 -.156 -.102 .005 Type III .001 .987 .012 .000 -.294 -.162 -.100 -.049 .029 Type IV .005 .985 .010 .000 -.250 -.153 -.097 -.046 .054 Type V .000 .000 .620 .380 .061 .175 .235 .286 .379 Type VI .000 .000 .995 .005 .028 .110 .171 .227 .328 Type VII .223 .777 .000 .000 -.334 -.246 -.184 -.137 -.051 Type VIII .996 .004 .000 .000 -.453 -.364 -.310 -.271 -.183 Sources: Labor Force Survey (1999-2002), Census of the Population (1999), INSEE, and General Transport Survey (2000), DREIF. Note: Statistics are computed on 1.000 simulations. The first four columns respectively provide the percentage of negative coefficients significant at the 5% level, negative not significant at 5%, positive not significant at 5%, and positive significant at 5%.

Graph 1. Sign and significance of the effects of living in a type-III neighborhood on labor-market transitions (under increasing degrees of endogeneity)

Transition: disappearing from the panel

0%10%20%30%40%50%60%70%80%90%

100%

[1/2,2] [1/3,3] [1/4,4] [1/5,5]

Odds-ratio interval

Perc

enta

ge s. >0ns. >0ns. <0s. <0

Transition: finding a job

0%10%20%30%40%50%60%70%80%90%

100%

[1/2,2] [1/3,3] [1/4,4] [1/5,5]

Odds-ratio interval

Perc

enta

ge s. >0ns. >0ns. <0s. <0

Transition: dropping out of the labour force

0%10%20%30%40%50%60%70%80%90%

100%

[1/2,2] [1/3,3] [1/4,4] [1/5,5]

Odds-ratio interval

Perc

enta

ge s. >0ns. >0ns. <0s. <0

s: significant at the 5% level. Ns: not significant at the 5% level.

Graph 2. Sign and significance of the effects of living in a type-V neighborhood on labor-market transitions (under increasing degrees of endogeneity)

Transition: disappearing from the panel

0%10%20%30%40%50%60%70%80%90%

100%

[1/2,2] [1/3,3] [1/4,4] [1/5,5]

Odds-ratio interval

Perc

enta

ge s. >0ns. >0ns. <0s. <0

Transition: finding a job

0%10%20%30%40%50%60%70%80%90%

100%

[1/2,2] [1/3,3] [1/4,4] [1/5,5]

Odds-ratio interval

Perc

enta

ge s. >0ns. >0ns. <0s. <0

Transition: dropping out of the labour force

0%10%20%30%40%50%60%70%80%90%

100%

[1/2,2] [1/3,3] [1/4,4] [1/5,5]

Odds-ratio interval

Perc

enta

ge s. >0ns. >0ns. <0s. <0

s: significant at the 5% level. Ns: not significant at the 5% level.